SESSION SU-M-1 : MOBILE ROBOTS AUTHOR(S) PAPER TITLE M. S. KIM, S. C. KANG, C. H. CHO, C. W. PARK, , C.-W. LEE (KIST, KOREA), C. H. LEE AND Y.-K. KWAK (KAIST, KOREA)

DEVELOPMENT OF A MOBILE ROBOT WITH DOUBLE TRACKS FOR HAZARDOUS ENVIRONMENT APPLICATIONS (ROBHAZ-DT)

K. NONAMI (CHIBA UNIVERSITY, JAPAN)

DEVELOPMENT OF AUTONOMOUS MINE DETECTION SIX-LEGGED WALKING ROBOT FOR HUMANITARIAN DEMINING

SESSION SU-M-2 : LEGGED ROBOTS AUTHOR(S) PAPER TITLE P. GONZALEZ DE SANTOS, E. GARCIA, J. ESTREMERA AND M.A. ARMADA (INSTITUTO DE AUTOMÁTICA INDUSTRIAL-CSIC, SPAIN)

SILO6: DESIGN AND CONFIGURATION OF A LEGGED ROBOT FOR HUMANITARIAN DEMINING

J.-C. HABUMUREMYI, I. DOROFTEI, Y. BAUDOIN (ROYAL MILITARY ACADEMY BRUSSEL, BELGIUM)

INTEREST OF WALKING ROBOTS IN HUMANITARIAN DEMINING PROJECT

SESSION SU-A-1 : DEMINING ROBOTS AUTHOR(S) PAPER TITLE Š. HAVLÍK (SLOVAK ACADEMY OF SCIENCES, SLOVAKIA)

MINE CLEARANCE ROBOTS

P. KOPACEK (VIENNA UNIVERSITY OF TECHNOLOGY ,AUSTRIA)

ROBOT SWARMS FOR DEMINING - VISION OR REALITY

A. T. DE ALMEIDA, L. MARQUES(UNIVERSITY OF COIMBRA, PORTUGAL), M. RACHKOV (MOSCOW STATE INDUSTRIAL UNIVERSITY, RUSSIA), V. GRADETSKY (RUSSIAN ACADEMY OF SCIENCE, RUSSIA)

ON-BOARD DEMINING MANIPULATOR

CH. GRAND, F. BEN AMAR, F. PLUMET AND PH. BIDAUD (UNIVERSITE DE PARIS VI, FRANCE)

SIMULATION AND CONTROL OF HIGH MOBILITY ROVERS FOR ROUGH TERRAINS EXPLORATION

SESSION SU-A-2 : MINE DETECTION I

AUTHOR(S) PAPER TITLE Y. CARON, P. MAKRIS AND N. VINCENT (UNIVERSITE DE TOURS, FRANCE)

COMPUTER VISION, AN HELP FOR DEMINING

G. DE CUBBER, H. SAHLI (VRIJE UNIVERSITEIT BRUSSEL, BELGIUM) AND H. P. E. COLON (ROYAL MILITARY ACADEMY, BRUSSEL, BELGIUM)

A COLOR CONSTANCY APPROACH FOR ILLUMINATION INVARIANT COLOR TARGET TRACKING

R. CHESNEY AND Y. DAS (CANADIAN CENTRE FOR MINE ACTION TECHNOLOGIES, CANADA)

TERRAIN ADAPTIVE SCANNING OF CONVENTIONAL MINE DETECTORS

SESSION MO-M-1 : MINE DETECTION II AUTHOR(S) PAPER TITLE O. DURAN, K. ALTHOEFER AND L. D. SENEVIRATNE (KING'S COLLEGE LONDON, UK)

AUTOMATED MINE DETECTION ALGORITHMS FOR SOILS WITH MINERAL CONTENT

SESSION MO-M-2 : TELEOPERATION AUTHOR(S) PAPER TITLE R. CHESNEY (CANADIAN CENTRE FOR MINE ACTION TECHNOLOGIES, CANADA)

AUGMENTED TELE-OPERATION

M. D. PENNY, S. COTTER, N. BEAGLEY, N. SMITH, AND K. WONG (QINETIQ LTD, UK)

A COMPARISON STUDY OF OPERATOR PERFORMANCE WITH THREE CONTROLLERS FOR A REMOTELY OPERATED VEHICLE

SESSION TU-M-1 : VARIOUS ASPECTS AUTHOR(S) PAPER TITLE M.-C. NGYUEN (TECHNICAL UNIVERSITY OF MUNICH, GERMANY)

GENERATING DISTANCE-INVARIANT OBJECT REPRESENTATIONS BY SUBSAMPLING OF IMAGES

T. WOJTARA AND K. NONAMI (CHIBA UNIVERSITY, JAPAN)

MINE CLEARANCE ROBOT HAND FOR A DEMINING ROBOT

P. KOPACEK (VIENNA UNIVERSITY OF TECHNOLOGY, AUSTRIA)

A "TOOL KIT" FOR DEMINING

Development of a Mobile Robot with Double Tracks for Hazardous Environment Applications (ROBHAZ-DT)

Munsang Kim, Sungchul Kang, Changhyun Cho, Changwoo Park, Chong-Won Lee, Cheonghee Lee* and Yoon-Keun Kwak*

Advanced Robotics Research Center, KIST, Sungbuk Ku , Seoul, Korea

*Department of Mechanical Engineering, KAIST Yoosung Ku , Teajon, Korea

Abstract In this paper, the design and the integration of the ROBHAZ-DT are introduced which is a newly developed mobile robot system with double tracks. It is designed to carry out military and civilian missions in various hazardous environments. The rotational passive mechanism equipped between the front and rear body makes the ROBHAZ-DT possible to have good adaptability to uneven terrain including stair. The passive adaptation mechanism reduces energy consumption in moving on uneven terrain as well as it offers simplicity in the design of the ROBHAZ-DT. Based on this new design concept, dynamic simulation was conducted to determine the significant parameters such as the optimal track size and the allowable attack angle. Also dynamic effects by vehicle turning are investigated to assess proper load torque. The ROBHAZ-DT system developed was successfully experimented in stair-climbing case. Keywords : mobile robot, track mechanism, hazardous environment, stair climbing, teleoperation 1. Introduction

In the 21st century, robots will take the place of human labor in many areas. In the near future they will perform various hazardous works like fire fighting, rescuing people, demining, suppressing terrorist outrage, and scouting enemy territory. To make use of robots in these various circumstances, robots should have the ability of passing through rough terrain such as steps. There are three types of moving mechanisms for this kind of robots in general; wheel type, track type and walking type mechanism. Robots with wheel mechanism are inferior to robots with track when they are to pass through rough terrain. Walking robots have complex structures so that they are usually difficult to control and slower in speed. In that sense, the track mechanism has advantages in high speed driving and mobility under severe conditions [1]. Contrary to these merits, it consumes much more energy than the other ones. Therefore it is needed to design a robot to overcome this drawback. Recent research trends are to develop track type robots with flexible configurations according to the ground conditions for more efficient adaptation to landform.

Kohler, et. al. suggested a moving mechanism with 4 flexible tracks [2] and Maeda, et. al. proposed

a moving mechanism that integrate the characteristics of robots with 4 wheels and track type robots [3]. Iwamoto and Yamamoto developed a moving mechanism that shows superior ability on going up steps with track that changes its configuration while moving [4]. Yoneda, et. al. suggested a track type moving mechanism that uses a track with a material having higher coefficient friction and wider contact area between the track and steps [5]. Schempf, et. al. suggested a robot with flexible track that has a similar structure to the robot proposed in this paper, but this robot must change the shape actively to adapt to the ground condition [6].

This paper suggests a track type whose front and rear tracks are to be rotated passively to adapt to the ground condition. Its superior ability was tested in uneven terrain driving like climbing stairs.

2. Design and structure of ROBHAZ–DT

In this section, we describe the design and simulation procedures that we have conducted to extract good design parameters.

2.1. Design of the track mechanism

The ROBHAZ–DT adopts a new track mechanism with flexible shape based on a unique link structure with passivity to pass through rough landform. Figure 1 shows the outline of the link type track mechanism applied to the ROBHAZ-DT. As shown in the figure, the link type track mechanism is composed of 4 tracks and 4 motors. Figure 1 (a) shows the side view of the mechanism. As the two tracks (�, �) are connected at the joint C passively, it can improve its adaptability to the circumstance without additional actuators.

Figure 2 shows the change of track configuration according to the driving direction of the motors. If both motors are driven in opposite direction as shown in Figure 2, each track can form an inverted triangle. With this configuration it can improve the energy efficiency in fast driving or turning phase by reducing the contact area against the ground.

2.2 Dynamic simulation of ROBHAZ–DT

Dynamic simulation was performed for the case

of climbing stair to find out the optimal design parameters. RecurDyn, a commercial multi-body dynamic simulation package [8] is used for this

purpose. In this section we will discuss about the general formulation of the dynamic model with kinematical constraints for the proposed double track mechanism [7].

(a) Side view

(b) Top view

Figure 1. Outline of a wheel mechanism

••••

Figure 2. Change of track shape according to the driving direction of the motors

2.2.1 Dynamic model

Motion of a body i is defined in general vector form as i

vie

TiiiC QQλqM

q+=+CC , i =1,..,n (1) (1)

where n is the number of bodies. iM is the mass matrix of body i,

iCq the constrained jacobian

matrix, and λ the vector of Lagrange multipliers. ieQ is the vector of external forces and i

vQ is the vector of centrifugal forces. iq is the generalized coordinate of the body i defined as [ ]Tiii θRq = (2) The vector iq has 6 dimensions. Ri represents a

transnational term (i.e., x, y and z) and θθθθi is a vector of rotational variable (i.e., θx, θy and θz). The number

of Lagrange multipliers is cn which represents the number of constraint equation. The constraint equation can be expressed in vector form as ( ) 0, =tqC (3) With partial differentiation of eq. (3) by q, we obtains

dQqCq =DD (4)

where dQ is defined as ( ) qCqqCCQ qqq DDD tttd 2−−−= (5)

while ( )( )t

tt

t

tt

∂∂∂=

∂∂∂=

qCC

CC

q2

2 (6)

Combining (1) and (4) into one equation, the whole system equation of motion can be derived as

+=

d

veT

QQQ

λq

0CCM

q

q �� (7)

where M is the system mass matrix.

2.2.2 Stair climbing simulation

The dynamic model used in the RecurDyn simulation has about 150 bodies, which represent small track link elements. Some important parameters for stair climbing simulation are listed in Table 1.

Figure 3 shows an example of fail situation while the ROBHAZ-DT climbs up stair. Points A and B in Figure 3 (a) represent edges of the stair. Slip occurs on the edges. It is found that the contact point B on body F should not be located around the attack region of the track, while body R contacts on point A. This result suggests that the longer body length would be better in climbing the stair with the given dimension.

Table 1 Simulation parameters

Track Link 0.33kg Track Body 25kg

Mass

Main Body 5kg Height 0.18m Stair Dimension Depth 0.24m

Friction Coefficient 0.8 Velocity 3m/min The simulation result shows that a proper

rotational travel range of the passive joint is needed to be determined. When the ROBHAZ-DT climbs the first step, there exists an excessive deflection in the positive direction (P) as depicted in Figure 3(b). The ROBHAZ-DT is apt to fold, if there exists no positive rotational travel limit at the passive joint. Figure 3(b) shows a situation of large deflection toward the negative direction. The Body F rotates toward the negative direction (N) through the passive joint. A, B and C are contact points and each arrow represents the direction of exerting forces to the track

body. Slip occurs on point C and the exerting forces on point A and B produce large torques to fold the vehicle toward the negative direction.

The mass distribution of the ROBHAZ-DT has a significant effect on stable climbing. By doing the dynamic simulation we can determine a proper mass distribution condition of the whole mechanism.

�

�

�

�

����

����

���

����

����

������

(a) Fail on two edges

����

����

�

�

�

�

�

(b) Rotational travel limit

Figure 3. Fail Conditions

2.3 Torque analysis for turning motion In turning motion, a tracked vehicle usually

requires a large torque because of the difference between actuating and moving direction of each track element. The purpose of this analysis is to find an effect of friction in turning situation and to obtain a guideline to select proper motor powers and reduction ratios for the required payload. This analysis begins with calculation of the exerted force on the track element.

In turning motion on plane surface, the behavior of the track elements of the ROBHAZ-DT is regarded as that of an usual single track (Figure 4 (a)). Figure 4 (b) shows an infinitesimal element of this track. Analysis was carried out for an extreme case in which the vehicle turns around the origin O. It is assumed that surface contact is uniform on each track module. In Figure 4(a), r is the position vector of the infinitesimal element. The t is the tangential

vector to the rotation of the infinitesimal element. L and W denote the length and the width of the single tracked vehicle, respectively. oF is the actuating force by motor and R is the exerting force on the infinitesimal element due to the contact and the friction.

� X

Yr

t

L

y

WoF

oF

θ

(a) Single tracked vehicle

LF

O

R

(b) Infinitesimal element Figure 4. Analysis for turning motion

We can derive force equilibrium equation for

the infinitesimal element of the track depicted in Figure 4(b) like tjF R

LFo −=∑ ˆ (8)

where L

MgR 2µ= , j is the unit vector of Y.

The total exerting moment on the track vehicle is ( )∑ ∫ ∑−

×== 2

22

L

LdyIT FrθDD (9)

where jir ˆˆ2

yW += .

The unit vector of r is

+= jir

r ˆˆ2

1 yWu

(10)

where ( ) 222 yW +=r .

In equation (10), 0=⋅ tru and ktr ˆ=×u are

satisfied. Let jit ˆˆyx tt += then

=

− 10

2

21y

x

tt

Wy

yW

r (11)

Thus the tangential vector is

+−= jir

t ˆ2

ˆ1 Wy (12)

Substituting equation (10) and (12) into Equation (9),

∫

∫∑

−

−

−=

−==

2

2

2

2

2

22

L

Lo

L

Lo

dyRWF

dyRL

WFIT

r

rθCC (13)

( )∫∫ −−+= 2

2

222

22

L

L

L

LdyyWRdyR r (14)

Let ( ) ( )222

tttt eeaeeWy−− −=−= then Equation

(14 ) can be described as

( )∫∫−

−+= 2

1

22

2

2 4t

t

ttL

LdteeaRdyrR (15)

where

+

+−= 1log2

1 WL

WLt ,

+

+= 1log2

2 WL

WLt .

Integrating the equation (13) gives

2

1

221

21

162 22

2 t

t

tto teeWRFWI

+−−= −θCC (16)

The equation (13) gives a turning criterion for the tracked vehicle such as

2

1

221

21

1622

t

t

tto tee

LMgWF

+−⟩ −µ (17)

Equation 17 can be rewritten as φµ⟩oF (18)

where 2

1

221

21

1622

t

t

tt teeL

MgW

+−= −φ .

The φ is invariant as long as dimensions and mass are not changed.

In summary, this result gives proper load torque information used for choosing a suitable motor. The turning load with the dimension described in Table 2 is greater than the starting load on 40° slope, while µ=0.8.

Table 2. Specification of ROBHAZ-DT External size (W × H × L) 484 × 319 × 720 mm

Weight (Battery included)

50 Kg

Nominal Velocity 0.86 m/sec (3.1Km/h) Max. Velocity 2 m/sec (7.2Km/h) Maximum angle for going up steps

40°

Passive joint limit +10 ~ -30

Power Ni-Cd rechargeable Battery

��������������� 1hrs

2.4. Structure of ROBHAZ–DT Figure 5 represents the appearance of

mechanical hardware of ROBHAZ–DT. The main body part is passively connected with the front and rear track parts. The main body is designed to be inclined according to the change of the relative angle between the front and rear track parts. A hydraulic damper is attached on each side of the main body and to absorb impacts from steps and rough landform. To detect unexpected obstacles 8 ultrasonic sensors are mounted at the body, whose information are to be used to make the operator control the robot easily. To control the robot in a remote area, a pan-tilt stereo camera that can monitor the surroundings of the remote area is used.

Figure 5. Appearance of the ROBHAZ–DT

3. Control Figure 6 shows the overall control architecture

of the ROBHAZ-DT. The control system has major features such as a force reflection, which enables an operator to feel the forces transmitted from the remote robot. By using this information from the ultrasonic sensors the control device can give the operator force sensation to effectively escape obstacles. The ultrasonic sensors attached on the front part of the robot form a pair with two sensors and face the ground in different angles so that they can detect not only obstacles like walls but also a ditch or cliff. Ultrasonic information used for force feedback is also used for escaping obstacles in a reactive way. As the emergent escape can be executed prior to the user’s input, primary safety can be guaranteed.

The three axes posture sensors mounted at the robot body enables the remote operator to monitor posture of the robot. If the robot is inclined over the allowable angle limit, it can retreat or rotate to a reverse direction.

Figure 6 is a block diagram of an integrated platform for the control system of the ROBHAZ-DT. The whole integrated system is composed of 1 PC and three sub-controllers, a remote PC, various

wireless modules and sensor modules. The PC exchanges information with the remote robot, controls to escape the obstacles as described in the previous section, performs the detection and modification of posture. Each sub-controller has a PIC microprocessor and controls two motors. The sub-controllers are connected to the PC through CAN (Control Area Network) bus and exchange the speed data of motors and various sensor information. This control structure enables a high responsive control by coordinating the sub controllers and non-real time control in the PC and distributing the calculation load of the system. The information exchange between the robot and remote operator is done through a wireless LAN. The robot sends various information of sensor to the PC and the PC commands the robot to move to a direction with a desired speed.

Figure 6. A block diagram of integrated control system of the robot

4. Experiment

The functional specification of the ROBHAZ-

DT is briefly described in table 2. Figure 7 shows the landform adaptability of the passive mechanism when the robot is going up stairs. Figure 7 (a) and (b) show the folded and unfolded states of the front and rear track part of the robot due to passivity. Figure 8 describes a movement on irregular landform that has about 30 degrees of slope in the roll and pitch direction.

5. Conclusion

In this work, a new track type mechanism was

proposed aiming at practical applications in hazardous environment. A shape configurable double track mechanism was designed to improve the driving ability under severe ground conditions like stair. Accomplishing a comprehensive dynamic simulation could optimize the dynamic property of the proposed mechanism. Some major functions like the force reflecting control scheme using ultrasonic

sensor signal were discussed. The experimented results show good mobile capability insensitive to ground condition and the dexterous remote control performance and prove the validity of our design idea on the passive mechanism.

(a) Positive deflection

(b) Negative deflection

Figure 7. Moving up and down the steps

Figure 8. Moving at irregular landform

References

[1] T. Iwamoto, H. Yamamoto, “Mechanical Design of Variable Configuration Tracked Vehicle,” J. of Mechanical Design, Vol. 112, 289-294, 1990.

[2] G. W. Kohler, M. Selig, M. Salaske, “Manipulator Vehicle of the Nuclear Emergency Brigade in the Federal Republic of Germany,” Proc. of 24th conf. On Remote System

Technology, 196-218, 1976. [3] Y. Maeda, S. Tsutani, S. Hagihara, “Prototype of

Multifunctional Robot Vehicle,” Proc. of Int. conf. on Advanced Robotics, 421-428, 1985.

[4] T. Iwamoto, H. Yamamoto, “Stairway Travel of a Mobile Robot with Terrain-Adaptable Crawler Mechanism,” J. of Robotic Systems, Vol. 2 No. 1, 125-134, 1985.

[5] K. Yoneda, Y. Ota, S. Hirose, “Development of Hi-Grip Crawler using a Deformation of Powder,” JRSJ Vol. 15 No. 8, 1188-1193, 1997.

[6] H. Schempf, E. Mutschler, C. Piepgras, J. Warwick, B. Chemel, S. Boehmke, W. Crowley, R. Fuchs, J. Guyot, “Pandora: Autonomous Urban Robotic Reconnaissance System,” Proc. of Int. conf. on Robotics and Automation 1999, Vol. 3, 2315 -2321, 1999.

[7] H. C. Lee, J. H. Choi, A. A. Shabana, “Spatial Dynamics of Multibody Tracked Vehicles Part II: Contact Forces and Simulation Results,” Vehicle System Dynamics, Vol. 29, 113-137, 1998

[8] “RecurDyn manual,” FunctionBay, Inc., 2000

Development of Autonomous Mine Detection Six-Legged Walking Robot for Humanitarian Demining

Kenzo Nonami*

Dept. of Electronics and Mechanical Engineering, Chiba University,

1-33 Yayoi-cho, Inage-ku,Chiba 263-8522, Japan nonami@meneth.tm.chiba-u.ac.jp

Abstract The present paper proposes two kind of mine detection robots which mean the six-legged walking robot with two manipulators based on the added stability, mobility, and functionality. One robot is COMET-II which is driven by electric power and not so big robot. The another robot is a full autonomous robot COMET-III which is driven by hydraulic power. COMET-III has a crawler and six legs and its weight is 900Kg. The improved this kind of robot will be engaged for mine detection job in Afghanistan soon.

Keywords: Humanitarian Demining Mine Detection Robot, Mine Clearance, Walking Robot, Autonomous Robot

1. INTRODUCTION Currently, more than 100 million anti-personnel mines are under the ground all over the world. Figure 1 shows the number of landmines in the most affected continents. These mines not only disturb the economic development of mine-buried nations, but also injure or kill more than 2000 people a month. As a result, the removal of landmines has become a global emergency. The current method of removing mines manually is costly and dangerous. Moreover, removal of all mines by this method would require several hundred years (it takes one thousand according to CMAC report based on Cambodian Mine Action Center Current Activities 1998), during which time, more mines might be buried in war zones. There are three kinds of demining strategies. The first is a human deminier based demining. The second is a mechanical equipment based demining like Fig.2. The third is a advanced robot based demining. Currently, the most demining approach is first type. Some limited mine field area are applied by the second type. The third type is now research oriented and expected in future’s demining approach. Figure 3 shows the robotics brush cutter which was developed in Demining Technology Workshop in Cambodia. This robot can work as cutting grass and bush by teleoperation. Figure 4 shows the four legged mine detection robot which was studied by CSIC-IAI, Spain in 1999. Also, the small and light eight legged, six-legged robots were made in FZI-S in 2001 for humanitarian demining.aim. Under this ultimate environment, a walking robot may be an effective and efficient means of detecting and removing mines while ensuring the safety of local residents and people engaged in the removal work. The

present paper proposes an improved six-legged walking robot COMET-II with two manipulators based on the added stability, mobility, and functionality that this platform offers. Also, the latest robot COMET-III which is a full autonomous robot is presented.

Fig.1 Landmines buried world map

Fig.2 Mechanical demining equipment

Number of Landmines in the Most Affected Continents

Asia Estimated more than 32 mill. landmines Afghanistan: 10 mill. Cambodia: 8-10 mill. China: 10 mill. Vietnam: 3.5 mill.

AfricaEstimated more than 44 mill. landmines Egypt: 15-20 mill. Angola: 9-10 mill. Mozambique: 3 mill.

EuropeEstimated more than 10 mill. landmines Bosnia: 3-6 mill. Croatia: 3 mill. Ukraine: 1 mill.

Middle EastEstimated more than 29 mill. landmines Iran: 16 mill. Irak: 7-10 mill.

Fig.3 Tempest for brush cutter

Fig.4 Four legged mine detection robot

Fig.5 Eight legged robot with outdoor

Fig.5 Eight legged outdoor and autonomy robo

Fig.6 Autonomous six-legged robot

First of all, we presented the mine detection six-legged robot COMET-I at MOVIC2000(1) and CLAWAR2000(2). The COMET-I has been tested on the out door many times. It is found that the COMET-I has a couple of problems. And then the improved walking robot COMET-II has been made(3). In particular�the new two manipulators were developed to attach at the front of the robot�The two manipulators are used for mine detection and grass cutting each other. One of two manipulators has a mixed sensor which means a metal detector and a radar sensor. And the other manipulator has a grass cutter. However, some new ideas were applied to realize a high speed performance and a little light weight. And, anti-personnel mines and anti-tank mines could be detected by the mixed sensor at the end of the manipulator. In the case of COMET-I, a robot should lower each leg onto the ground safely and stably without stepping on a mine. It took much time to walk. However, COMET-II has faster locomotion speed and the speed is 150m/h with searching mines. The mine detecting area is about 300 m2 per hour. The COMET-II has ten times speed comparing with COMET-I. When detecting a mine and a UXO's by such sensors, the GPS based mapping has been made automatically. And then, the avoidance walking control algorithm guards COMET-II against anti-personnel mines and UXO's. This is a very important speciality of COMET-II. Also, we have already verified the efficiency by means of walking experiments on rough surface like rough terrain using neuro-based hybrid position and force control. The nonlinear control based on neural network(4)(5)(6) has been applied for motion control of COMET-II. In our control strategy, the central hierarchical neural network (HNN) controller, which is used to make the robot realize a desired motion. Consequently, the autonomous locomotion of the robot was realized. The experiments were performed to verify the effectiveness of the proposed control strategy. The latest mine detection robot COMET-III has been built. This robot is scaled up from COMET-II. The total weight is about 900Kg, the size is 4m long, 2.5m wide, 0.8m high. COMET-III has 40 litter gasoline tank to continuously work for 8 hours and 700cc gasoline engine like automobile engine to generate DC power supply. The driving force is based on hydraulic power with 14 MPa high pressure.

2. OUTLINE OF THE TELEOPERATED MINE DETECTION WALKING ROBOT COMET-II In the present study, a six-legged walking robot with each leg having three-degree-of-freedom (3DOF) was used to

ensure more stable walking on rough ground. The robot has six three-joint legs, each consisting of a shoulder, a thigh, and a shin. The leg mechanism uses a parallel link for the thigh and the ball joint of the parallel link turns the shin. The necessary power received from the power generator using gasoline for long time outdoor operation and from an external power supply for indoor operation. The total weight is about 100 kgf, the width is about 1400 mm, and the only body height is about 600 mm. Figure 7 shows the overview of COMET-II. The most speciality of COMET-II has two manipulators with the mixed sensor using metal detector and GPR at the tip of the right manipulator and the colour spray gun at the tip of the left manipulator. COMET-II has been equipped with two computers based on RT-Linux, 32 channels D/A and 64 channels A/D converters, 20GB hard disk for autonomous locomotion control, for external recognition and radio transmission/reception of data, and for mine detection and radio transmission/reception

Fig. 7 Overview of COMET-II of data. High-level motional control and external recognition will be assigned to an external high-speed host computer based on teleoperation. As another equipped devices, a visible-light camera, an infrared camera, an attitude control sensor, two radio transmission /reception antennas for image data, two telecommunication antennas for control commands, six force sensors, twenty four potentiometers, a ground penetrating radar (GPR) sensor and a GPS system with 1cm precision are equipped. In the case of COMET-I, a metal detector and an optical proximity sensor were attached to the foot of each leg to prevent the robot from stepping on a mine during walking on a mine field, and compliance control protects the robot legs from shocks. From this protection algorithm of robot against anti-personnel mines, COMET-I was so slow locomotion

speed. Therefore, the COMET-II does not have such devises anymore. Fist of all, the mixed sensor at the tip of the right manipulator looks for mines, metals and UXO’s moving right and left. And then, the danger and safety area are classified by GPS system with 1cm precision. These data are recorded on the computer. The COMET-II can avoid the dangerous area using such GPS based mapping. The mine detection speed of the COMET-II becomes so high speed which is ten times than COMET-I. 2.1 Mine Detection Method Let us explain how to detect mines or mine like objects. Figure 8 shows the mixed sensor head. ECT stands for the eddy current tester as metal detector and GPR means the ground penetrating radar. So, metals are detected by ECT and the objects ore than some size including plastic mine are detected by GPR.

Fig.8 Mixed mine detection sensor (ECT and GPR)

Figure 9 shows the summary how to recognize mines. Even plastic mine has a little bit metal. ECT has very high sensitivity to detect the such tiny metal against plastic mines. GPR should be tuned as to detect more than anti personnel mine like size. After that, the detection efficiency goes up. GPR has two kind of zone which means the shallow zone from o cm to 25 cm deep and the deep zone from 25 cm to 80 cm deep. Anti personnel mines are detected in shallow zone and anti tank mines are detected in deep zone. In the case of anti tank mines, it is made of full metal and 40cm in diameter. Therefore, it is very easy to detect them by ECT and GPR. Figure 10 shows the mine detection ratio. We have actual mines which were impoted from Cambodia. These mines were buried under the ground in Cambodia and CMAC detected them and disposed. So, we have been carrying out the mine detection test using these actual anti-personnel mines and anti-tank mines. From these results, it is found that anti tank mines are completely de-

1.�ECT+GPRshallow= Anti personnel mine or UXO’s(Possibility of mine(APM) like objects is so high)

2.�ECT = Fragment or noise(Size is small)

3.�GPRshallow= Stone or Plastic based mine(Size is not small)

4.�ECT+GPRdeep= Anti tank mine or UXO’s(Size is big with deep area and metal response)

How to recognize mine ?

Fig. 9 How to recognize anti personnel mines

98%NOMZ-2B

100%Anti-

TANK-Mine

86%PMN2

90%PMN

Detection ratioResponse signalPhoto

Mine detection ratio

Fig.10 Mine detection ratio against several mines tected by this mixed sensor. It is not easy to detect anti personnel mines with 100% reliability. Now, we have been studying to increase the detection ratio using neural network learning and pattern matching after database stock. We hope it will be close to 100% for detection of anti personnel mines. 2.2 Mine Avoidance Walking Control COMET-II has a special mission to make mine buried map. After that the mine clearance vehicles follow the COMET-II to dispose and clear anti personnel mines avoiding the dangerous spots like anti tank mines and UXO’s. Also, COMET-II has to guards against anti personnel mines. Therefore, we developed the mine avoidance walking control algorithm. Figure 11 shows the concept of mine avoidance walking control. At first, mine buried spots have been already detected and the robot can recognizes such GPS based mine buried spots. So, X0 is already given and the virtual mechanical components are assumed as Fig.5. Let us define the following equations (1)

Fig.11 Mine avoidance control using impedance control

where,

If the leg enters into a virtual sphere, the virtual external force is defined using non contact impedance as follows:

(2) Using this virtual external force, the control input should be modified at first step, and the next step the leg trajec- tory will be also modified. Figure 12 shows the such block diagram as mine avoidance walking control. Figure 13 shows the one example of experimental results. It is found that the proposed virtual impedance control method is a successful trajectories avoiding mine buried area.

Fig.12 Block diagram of mine avoidance walking control

Fig.13 Foot trajectory avoiding mine buried area

2.3 Obstacle Avoidance Walking Control In order to realize a full autonomous mine detection robot, we apply laser range finder to measure distance between the robot and the obstacle. In this case, we define the cost function as follows:

∫= LdxxuJ )( (3)

where )(xu is the danger index of collision against obstacles each position, L is the walking path and J is the cost function to minimize in order to avoid a collision. We difine the following function.

)()()( xExxu η= (4) where, )(xη is the inverse number of the distance from the robot to the obstacle , )(xE is the estimated error from the exact position of the robot. So, the )(xu of Eq.(4) should be smaller than 1 to avoid the collision. Fig.

Fig.14 Obstacle avoidance walking control against three obstacles

2.4 Neuro Based Position and Force Hybrid Control Figure 15 shows the control system for each leg. The control system for the entire robot is represented by six block diagrams shown in Fig.15, because the six legs have identical control systems. The control system has three components. The first component is a position control system with a PD feedback controller for each joint, and the second component is a force control system using a feedforward controller and a PD feedback controller. The first component and the second component constitute a conventional position and force hybrid control system. The third component, which is designed on the basis of neuro learning for attaining ideal control performance without use of a correct mathematical model of the leg, is a position and force hybrid control system that uses an inverse-function-type centralized control compensator for the three joints of one leg. The third component and the hierarchical control structure shown in Fig.15 are the concepts that are newly proposed in our study. The conventional hybrid postion/force control torque is given as follows: τ(t) = τp(t) +τf(t)

= τp(t) +τfb(t) +τff(t) = Kppθei(t) + Kpd θ

�ei(t) + Kfp τe(t) + Kfd + τ

�e(t) + Jf T fri

= Kppθei(t) + Kpd θ�

ei(t) + Jf ( Kfp fei(t) + Kfd f�ei(t)

+ T fri ) (5) This force control rule is based on the proposal of Raibert and Craig[7]. Equation (5) expresses the control input of conventional position and force hybrid control. However, because the four feedback gain matrices Kpp, Kpd, Kfp, and Kfd in Eq.(5) are obtained by rule of thumb and trial and error, good control performance cannot be realized. In this study, we propose a new position and force hybrid control system that is based on a neural network, which can realize good control performance without a correct Fig.15 Control system with neuro based position/force

hybrid controller mathematical model of the leg.

Using the hybrid control input τp + τfb, the neuro controller can determine the actual control input by feedback error learning. Eq.(6) expresses the desired control input of the neuro controller for each motor.

τd(t) = VjkT yj ( Wij

T xi ) + Kppθei(t) + Kpd θ�

ei(t) + Jf ( Kfp fei(t) + Kfd f

�ei(t) ) (6)

The input layer xi ( i=1…6 ) consists of the position and velocity of reference trajectories of the three joints of one leg. The output layer zk ( k=1…3 ) consists of the input torqueτn for three motors of the one leg as determined from inverse dynamics of the neuro compensator.

The relations among units of the respective layers are as follows:

xi(K) = θr1 (K), θ�

r1 (K), …, θrk (K), θ�

rk (K) (7)

yj(K) =

∑i

iij KxKWf )()( (8)

zk(K) =τnk(K) =

∑

jjjk KyKVf )()( (9)

i=1…6, j=1…12, k=1…3 where, the sigmoidal function is defined as follows:

( ) ( ) 11

2)( )()( −+

= +− KbKueKuf (10)

where, b(K) is the item for improving learning ability. Figures 16 shows the force target values (dotted lines)

for legs 1, responses of neuro-based position and force hybrid control (bold solid lines), responses of conventional position and force hybrid control (broken lines), and responses of solely position control (solid line). Fig.16(a) shows that the response of neuro based position and force hybrid control most closely approximates the force target value, followed by response of conventional position and force hybrid control, and response of solely position control. Tracing performance for force target value of neuro-based position and force hybrid control exceeds 92�, whereas that of conventional position and force hybrid control reaches only 50�. Of course, because tracing performance of conventional hybrid control depends on feedback gain matrices, performance is improved if proper gain can be obtained, at the expense of a great amount of time. However, in practice 3 unknown gains exist on the diagonal of each of the four gain matrices Kpp, Kpd, Kfp, and Kfd in Eq.(5). It is very difficult to determine the twelve unknown gains. Therefore, the new algorithm proposed in this study is required for the multi-legged walking robot. Fig.16(b) shows the control input corresponding to (a), in comparison with conventional hybrid control; in neuro-based hybrid control the

Fig.16 Experimental result of Leg1

vibration of control input is almost zero, and the control input itself is quite small. In other words, Fig.16 shows that control input is very small and that control performance is excellent. 3. OUTLINE OF THE AUTONOUMUS MINE DETECTION CRALER-LEG ROBOT COMET-III The latest mine detection robot COMET-III has been built. Figures 17 and 18 show the photos of COMET-III. This robot is scaled up from COMET-II. The total weight is about 900Kg, the size is 4m long, 2.5m wide, 0.8m high. COMET-III has 20 litter gasoline. Figure 19 shows the final version of COMET-III.

Fig.17 Fron view of COMET-III

Fig.18 Top view of COMET-III

Fig.19 Final version of mine detection robot COMET-III tank to continuously work for 6 hours and 650cc gasoline engine like automobile engine to generate DC power supply shown in Table 1. The driving force is based on hydraulic power with 14 MPa high pressure. The locomotion speeds by walking mode of legs and running mode of crawler are about 600m and 3km per hour respectively. From this walking speed , the mine

Table 1 Hardware Specification of COMET-III Hydraulic Pump 140kg/�, 17L/min Engine 653cc, Max Output 22ps/3600rpm

Max Torque 4.65kgf/m Material Aluminium Alloy, SUS304 Crawler Rubber Crawler Hydraulic Tank 40L Gasoline Tank 20L Control Valve 26 Thigh Cylinder φ25×250st, Max Speed 280mm/s Shin Cylinder φ30×175st, Max Speed 300mm/s Leg Turn Max Angle ±80°,69°/s Weight Frame+Crawler+Engine+Hydralic

Unit=468kg, Legs=32kg×6

Mine BuriedElectronicMappingControl

Multitask cooperativecontrol

Hierarchy supervisory controlHost

computer

Mine Detection

SensorControl

ReferenceTrajectoryFollowingControl

Color PaintMarking Control

ManipulatorHand

Control

Mine AvoidanceWalking Control

Obstacle AvoidanceWalking Control

CrawlerRunning

TrajectoryControl

Vision BasedEnvironmentRecoginition

Control

Fig.20 Hierarchy supervisory control of COMET-III detection speed becomes 1800 �/h. The robot will be able to climb up the slope with 30 degrees using crawler and legs. The payload will be about 300kgf. The COMET-III has a hierarchy supervisory control system shown in Fig.20. Also, the multitask cooperative control system is applied by SH4 computer systems. The embedding computer system is communicated with the host computer system simultaneously. The COMET-III has the mixed array mine detector at the right arm and the grass cutter at the left arm, and is almost corresponds to a real machine to mine detection robot in actual mine field. This robot is waterproof and can be working for not only day time but also night time, also all season.

4. CONCLUSIONS We propose a full autonomous mine detection robot. After a field test during one or two years, the proposed improved robots will work at mine detection in Afghanistan.

REFERENCES [1] K.Nonami, Dynamics and Control of Six Legged Walking Robot for Mine Detection, Proceedings of the Fifth International Conference on Motion and Vibration Control (MOVIC2000), pp.31-38, December 4-8, 2000, Sydney [2] K.Nonami, et al, Humanitarian Mine Detection Six-Legged Walking Robot, Proceedings of the Third International Conference on Climbing and Walking Robots (CLAWAR2000), pp.861-868, October 2-4, 2000, Madrid [3] K. Nonami, Q. Huang :Humanitarian Mine Detection Six-Legged Walking Robot COMET-II with Two Manipulators, Proc. of 4th International Conference on Climbing and Walking Robots , (CLAWAR2001), pp.989-996, 2001, Karlsruhe, [4] M. Kawato : Feedback-Error-Learning Neural Network for Supervised Motor Learning, Advanced Neural Computers, R. Eckmiller(Editor), pp.365-372, 1990. [5] J. Kodjabachian, J.A. Meyer : Evolution and Development of Nerural Controllers for Locomotion, Gradient-Following, and Obstacle-Avoidance in Artificial Insects, IEEE Transactions on Neural Network, 9-5, 1998. [6] K. Koyama and K. Nonami : Dynamic Walking Control of Quadruped Locomotion Robot Using Neural Network, Trans. of the Japan Society of Mechanical Engineers, C, 62(596), pp.1519-1526, 1996. (in Japanese) [7] M. H. Raibert and J. J. Craig : Hybrid Position/Force Control of Manipulators, ASME Journal of Dynamic Systems, Measurement, and Control, 103-2, pp.126-133., 1981.

Appedix

Figure 21 shows the illustration of humanitarian demining using autonomous mine detection robot, full autonomous radio-controlled small mine detection helicopter and mechanical clearance robots. The mine detection helicopter will fly very low height like 1m or 1.5 m over the ground and it has same array mixed sensor which consists of metal detector and GPR. The helicopter hangs such sensor using some arms to keep constant gap between ground surface and a sensor head. The mine detection robot and the mine detection helicopter will be able to recognize safety area and dangerous area. In particular, an anti-tank mines and UXO’s are so much dangerous objects for mechanical clearance robots. After marking and mapping all mines and UXO’s, unmanned mechanical clearance robots will dispose only anti-personnel mines avoiding anti-tank mines and UXO’s. The mine field has so many situations which mean dry and hard soil area, wet or soft soil area, bush area, small rock area, and so on. So, we have to prepare many kinds of mechanical clearance robots and the mine detection robot will conduct which machine will be best fit to clear some mines. All data will be sent to a host computer in one box car as control center and all data will be managed. These data will be transmitted to Japanese control center via an artificial satellite with real time two-way communication.

Fig.21 Illustration of humanitarian demining using mine detection robot and mechanical

clearance equipments in Afghanistan (This illustration was provided by Graphic Science Magazine “Newton”, August, 2002, Japan)

SILO6: DESIGN AND CONFIGURATION OF A LEGGED ROBOT FORHUMANITARIAN DEMINING

P. Gonzalez de Santos, E. Garcia, J. Estremera and M.A. Armada

Instituto de Automática Industrial-CSICCtra. Campo Real, Km. 0,200- La Poveda28500 Arganda del Rey, Madrid, Spain.

Abstract

Removal of antipersonnel landmines is a worldwideproblem that requires the use of new technologiessuch as improved sensors and efficient mobile robots.This paper is focused in studying the real possibilitiesof using walking machines as sensor carriers todetect and locate antipersonnel landmines in anefficient and effective way. This paper describes themain features of a legged robot endowed with amanipulator to accomplish with the detection andlocation of landmines. Sensor and locators are brieflypresented and some design features are advanced.

Keywords: Walking robots, legged robots,humanitarian demining, navigation, robot location.

1. Introduction

Detection and removal of antipersonnel landmines is,at the present time, a serious problem of political,economical, environmental and humanitariandimension. There is a common interest in solvingthis problem and solutions are being sought inseveral engineering fields.The finest solution, but maybe not the quickest, is toapply fully automatic systems to this relevant task.However, this solution seems to be still very far awayfrom succeeding. Efficient sensors, detectors andpositioning systems are, first of all, required todetect, locate and, if possible, identify the mines.Then, appropriate vehicles will be of paramountimportance to carry these sensors over the infestedfields keeping the human operators as far as possiblefor safety reasons. Fully automated systems arecomplex to achieve and an intermediate solutionmight be teleoperation and human-machinecollaboration in the control loop, which is beingknown as collaborative control.Any potential vehicle, in principle, can carry sensorsover an infested field: wheeled, tracked and evenlegged vehicles can accomplish demining tasks witheffectiveness. Wheeled robots are the easiest andcheapest; tracked robots are very good to move inalmost all kinds of terrain; but legged robots alsoexhibit potential interesting advantages for thisactivity. For instance:

• Legged robots only require a finite number ofground contact points decreasing the likelihoodof stepping on an antipersonnel mine.

• Wheels and tracks describe a continuous path,whilst legs only need to stand on discrete pointsalong the path. This would allow locating allpotential alarms in a field before starting thedeactivation task.

• The inherent omnidirectionality of legged robotsis also a great advantage for changing thesteering without performing forward/backwardmaneuvers.

• Legged robots can negotiate irregular terrainmaintaining the body always leveled. This isimportant for carrying on-board sensors andequipment that need to be leveled whilemeasuring.

• Walking on a slope with the body leveled is aneasy task for legged robots without jeopardizingtheir stability.

• Legged robots can walk over loose terrain suchas sand, and legs endowed with the proper forcesensors can identify the stepped terrain toprevent slippage.

• A legged robot provides additional motionsalong the x, y and z components and even bodyrotations without changing its footprints.

The idea of using legged mechanisms forhumanitarian demining has been developed for, atleast, the last five years and some prototypes havebeen already tested. TITAN VIII [7], AMRU-2 [1]and RIMHO2 [3] are some more examples ofwalking robots used as testbeds for humanitariandemining tasks. COMET-1 is maybe the first leggedrobot developed on purpose for demining [10].These four robots are based on insect configurations,but there are also different legged robotconfigurations, such as sliding frame systems, beingtested as humanitarian demining robots [6], [9].The IAI-CSIC legged-robot working team accredits alarge experience in the development of walkingrobots [2], [3], [4], [5]. Since 1999 it has beenworking in the application of legged robot fordetecting and locating unexploded ordnance as a veryimportant potential task for this kind of locomotion.This team, funded by CICYT (Spanish Ministry ofScience and Technology) is currently developing a

legged robot for efficient detection and location ofantipersonnel land mines (DYLEMA project). Themain aim of this project is to develop a leggedprototype to integrate relevant technologies in thefields of legged locomotion and sensor systems inorder to identify the needs of this integration forhumanitarian-demining activities. This paperpresents some preliminary ideas about: the robotconfiguration, sensors, locators, control system andcontrol strategies.

2. System description and main requirements

The DYLEMA project's main aim is to develop arover able to detect and locate antipersonnel landmines. For this purpose, it is required, at least, thefollowing subsystems illustrated in Figure 1:1. Detector. In this project the detection system is

reduced to just a single metal detector.2. Scanner. When using a single sensor it could be

necessary to use a scanning system to swap largeareas. The most efficient scanning system for asingle sensor is a robot manipulator. In this case,a manipulator of 5 DOF will be developed forthis task.

3. Locator . After detecting a suspect object thesystem has to mark the exact location in adatabase for a posterior analysis anddeactivation. Differential Global PositioningSystem (DGPS) seems to be the suitabletechnology. However, it exhibits some lacks: itis extremely expensive, prone to malfunctions inareas covered with natural objects (trees, largerocks, etc) and requires additional equipment(one more antenna) for getting accuracy. Anaccuracy of about ±2 centimeters is consideredas sufficient for locating land mines.

4. Mobile robot . A mobile platform to carry thedifferent subsystems across the infected field isof a vital importance. This platform is basedon a legged robot because of the advantagesmentioned above. The following requirements

are the starting point for configuring thewalking robot:

• The legged robot will be based on a hexapodconfiguration. Hexapods can, theoretically,achieve higher speed than quadrupeds.

• The legged robot should be lightweight enoughto be handled by two grown persons. Thisrequirement is important to rescue the robotfrom technical o logistic problems.

• The robot should be autonomous from theenergy point of view.

• The robot should be semi autonomous from thecontrol point of view. Thus, a remote operatorshould be in the loop to control the systemthrough teleoperation and collaborative control.

The robot is being configured to optimize the powerconsumption as well as the mobility and stability.These are antagonist conditions that are beingbalanced by a detailed design.5. Controller. The global control system will be

distributed in two main computers: on-boardcomputer and operator station. The on-boardcomputer is in charge of controlling andcoordinating the manipulator and leg joints,communication with the DGPS and detector aswell as communication with the operator stationvia radio ethernet.

6. Power Supply. DC batteries provide lowerweight than fuel generators, but also lessautonomy . Nevertheless, fuel cell technology isgrowing quickly and promises to be the bestoption in the near future.

Hence, the SILO6 walking robot is to be configuredas a six-legged-autonomous robot carrying amanipulator. The system will be controlled throughteleoperation and collaborative techniques .Following sections give a global view to the systemconfiguration focused on the leg and bodyconfiguration, sensor system and some controlleraspects.

3. Detection and location of land mines

This section describes the commercial systems usedin the DYLEMA project: detector and locator.Features are commented briefly and a smalljustification of the selection is provided.

3.1. Detector

There are different technologies for detecting mines .The simplest one consists in a metal detector device.These sensors are simple, lightweight and easy ofuse. However, they only detect mines with somemetallic parts and become inefficient with non-metallic mines (plastic mines). In such a case,another type of sensor such as those based on GroundPenetrating Radar (GPR), chemical sensors orartificial noses must be used.DYLEMA project is devoted with the development

Figure 1: DYLEMA Global system

Detector

Radio ethernetaerial

DGPSantenna

Walking robot

Operator stationPowersupply

Scanner(manipulator)

Onboardcomputer

of mobile robotic techniques for identification andlocation of land mines . No sensor development isunder the scope of the project. Therefore, a metaldetector seems to be the simplest selection for ademining sensor, just to help in the detection andlocation of potential alarms . After detection of asuspect object, its location must be marked in thesystem database for further analysis and possibledeactivation.For DYLEMA project purposes the commercial minedetecting set Schiebel AN-19/2 is used. This detectoris in service in the US Army as well as in severalNATO countries . It has been designed to detecttypical mines with a very small proportion of metalcontent.

3.2. Locator

Marking the position of any suspect object ismandatory in demining tasks . In an automated orsemi-automated system a computer database seemsto be the most efficient manner for keeping a recordof alarms . But, first of all, it is necessary to locateaccurately the potential alarms . GPS technique is agood candidate: simple to use and accurate enough.We have planned to study alternative solutions, but aGPS system is going to be considered as the firstsolution as well as the system for calibratingoncoming new systems . DYLEMA project requireslocating alarms with an accuracy of about ±2centimeters . Therefore, the Real Time Kinematics(RTK) technique will be used because it can providethe expected accuracy. It requires using an additionalGPS antenna that will be placed at the operatorstation (see Figure 1). With these preliminaryspecifications the DGPS 4700, manufactured byTRIMBLE, is an adequate candidate for ourapplication.

4. Scanner

DYLEMA project will use a metal detector, which isa device that senses small areas . Therefore, a scannerdevice able to carry the sensor across larger areas isnecessary . The easiest system could be a

manipulator developed on purpose. A 5-DOFmanipulator has been chosen as the most appropriatestructure to accomplish the task.The manipulator is designed just to carry the selecteddetector; therefore, the design is optimized forcarrying the detector load. First, the load is balancedto move the detector ±45º in its pitch and roll wristaxes with the less torque. This is accomplished byplacing the detector is a configuration in which notorque is required in the normal position (detectorparallel to the ground). Regarding the manipulator, aRRR arm configuration is good enough for thisapplication. Figure 2 shows a preliminary drawing ofthe scanner configuration. Manipulator link lengthswill depend on the walking robot dimensions (bodyheight and leg span).

5. Walking robot configuration

Walking robots are intrinsically slow machines and itis well known that, from the theoretical point of view,machine speed depends on the number of legs.Therefore, a hexapod can achieve higher speed than aquadruped and a hexapod achieves its higher speedwhen using a wave gait with a duty factor β = 1/2,that is, using alternating tripods [11]. Althoughstability is not improved very much when usingalternating tripods, a hexapod configuration has beenchosen just to try to increase the machine’s speed.Hence, the mobile platform will exhibit an insect-likeconfiguration.

5.1. Body structure

The body of a walking robot has the main tasks ofsupporting legs and accommodating subsystems .Therefore, is must be big enough to contain therequired subsystems: on-board computer, electronics,drivers, GPS, batteries, etc. Preliminary volumes ofthese subsystems define the volume of the body.Alternating tripods means that two non-adjacent legsof one side and the central leg of the opposite sidealways support the robot. That means that for agiven position of feet the central leg in support phaseis carrying about half the robot’s weight, while thetwo collateral legs in support phase are carryingabout ¼ of the robot’s weight. This is especiallysignificant in traditional hexapod configurations inwhich legs are placed at the same distance from thelongitudinal axis of the body. If the robot has similarlegs, then the non-central legs will be overdimensioned and to optimize the mechanism thecentral leg design should differ from the rest of legs.However, using just one leg design has manyadvantages in reference with design cost,replacements, modularity and so on.

Figure 2. Scanner (manipulator arm)

±45º

Roll and pitch joints

Elbow joint

Waist and shoulder joints(differential system)

Pitch motor

Roll motor

Elbow motor

A satisfactory force distribution and homogenizationof the system can be accomplished by displacing thecentral leg a little bit from the longitudinal body axis .In this case the central legs support less weight andthe corner legs increase their contribution insupporting the body. This effect is illustrated inFigure 3, which represents a legged robot supportedon three legs. The equilibrium equation that balanceforces and moments is given by [8]:

1 5 4 1

1 5 4 5

4

00

1 1 1

x x x Fy y y F

F W

=

(1)

The condition for sharing homogeneously the weightof the robot among the supporting legs is:

4

0 / 3 0/ 3 0

1 1 1 / 3

X X WY Y Y W

W W

− − =

(2)

Rows 1 and 3 are always satisfied and row 2 issatisfied if:

42Y Y= (3)This last condition produces an unusual configurationthat looks not very suitable for our application. Inany case, the further the central foot is from thelongitudinal body axis the more homogeneous forcedistribution results.The adopted solution has been to choose equal legsand locate the central legs in an advanced positionwith reference to the longitudinal body axis . Thefinal leg location has been a compromise betweenbody shape and leg positions.The Yobotics Simulation Construction Set has beenused to compute leg torques . These torques are thestarting point for actuator selection. Figure 4 showsa graphic simulation of the walking robot endowedwith a manipulator (scanner). The simulation hasbeen computed using the leg structure anddimensions presented in the following section.

5.2. Leg structure

Walking robots need leg configurations that providejust contact points with the ground. Therefore, a 3-DOF device is enough to accomplish motion. Legshave to be designed to be lightweight mechanisms

and have to support the robot's weight. Therefore,the load carried by each leg is very high and can besupported with the leg in a different configuration. Amammal configuration is the most efficient legconfiguration from the energy point of view (lesstorques are demanded). However, it is not veryefficient if we analyze stability. Insect-like legs seemto be more efficient in this case; nevertheless, powerconsumption increases extraordinarily with thisconfiguration. The idea is to provide a legconfiguration able to accomplish with both stabilityand energy efficiency (very important in outdoormobile robots); therefore, a leg able to get mammaland insect configuration is developed. The startingpoint is to consider the torques the robot has toprovided in an insect configuration (the worst case).These torques, for the selected body configuration,have been computed through simulation. Anadequate way of decreasing motor size is by usingactuators in parallel; that is, actuators are placed insuch a way that two actuators work at the same timeto accomplish motion in one joint. Thisconfiguration gives the benefits of using smallmotors . Therefore, a differential driving mechanismwill be used for joint 2 and 3. Figure 5 shows apreliminary design of the leg.Feet can be designed in two basic configurations: aball fixed to the ankle or a flat sole with somearticulated passive joints . The first one is thesimplest and can be sufficient for our application ifthe radius of the ball is big enough to preventsinking.

6. Control system

The control system is distributed between theoperator station and the on-board controller. Both ofthem consist in PC-based computers (see Figure 1).The operator station will run under Windows NToperating system and the on-board controller (robot'scontroller) will run under QNX, a real-timemultitasking operating system. Communicationbetween operator station and on-board computer willbe performed by radio ethernet. The followingsections illustrate the main hardware and softwareaspects.

Figure 3. Force distribution for a tripodconfiguration Figure 4. SILO6 sketch in simulation.

F4(0,-Y4) F1

(X,Y)F5

(-X,Y)

z yx

6.1 Hardware architecture

The on-board controller is a distributed hierarchicalsystem composed of a PC-based computer, a data-acquisition board and eight three-axis control boardsbased on the LM629 microcontrollers, interconnectedthrough an ISA bus. The LM629 microcontrollersinclude digital PID filters provided with a trajectorygenerator used to execute closed-loop control forposition and velocity in each joint. Everymicrocontroller commands a DC motor-joint driverbased on the PWM technique. An analogue data-acquisition board is used to acquire sensorial datafrom the different external equipment (sensors,locators, etc.). A radio ethernet card is provided fornetwork communication with the operator station.Additional electronic cards for interfacing thedetector and the DGPS (via RS-232) are alsoprovided. A general diagram of the SILO6 hardwarearchitecture is shown in Figure 6. The operatorstation hardware is based on a standard computer.

6.2 Software architecture

The on-board computer is in charge of the walkingrobot gait and trajectory generation, manipulatorcontrol, signal processing and communications, aswell as coordination of the microcontrollers. Thesetasks are distributed in a software architecture thatconsists of layers developed on a bottom-up basis.These layers can be mainly divided into:• Hardware interfaces: These layers contain the

software drivers for both walking robot andmanipulator.

• Axis control layers: These layers perform thecontrol of individual robot joints for bothwalking robot and manipulator. Individual jointsare controlled through a dedicated micro-controlled, which runs a PID control algorithm.

• Leg control: This layer is in charge ofcoordinating all three joints in a leg to performcoordinated motions.

• Leg kinematics: This layer contains the directand inverse kinematic functions of a leg.

• Trajectory control: This module is in charge ofcoordinating the simultaneous motion of all fourlegs to perform straight-line or circular motions.

• Stability module: This layer determines whethera given foot position configuration is stable ornot.

• Gait generator: This layer generates the sequenceof leg lifting and foot placement to move therobot in a stable manner. The stability moduleguarantees static stability. The SILO6 gaitgenerator will be based on three gaits: a tripodgait, a spinning gait and a turning gait.

• Communications: This layer states thecommunication with the operator interfacethrough radio ethernet via TCP/IP protocol.

• Manipulator kinematics: This layer is in chargeof solving the manipulator kinematics.

• Equipment and sensor data acquisition layer:provides the required interfaces with externalequipment.

Figure 7 diagrams the different software modules andtheir interconnections.

7. Conclusions

Detection and location of antipersonnel landmines isbeing mainly carried out by human operator handlingmanual equipment. Robotization of this activity cangive so many benefits to the human community inmany countries . There is a common interest insolving this problem and solutions are being soughtin several engineering fields.New sensors are being required to detect land minesefficiently, but current sensors can be carried by

Figure 5. SILO6 leg configuration. Figure 6. SILO6 hardware architecture.

DGPSantenna

Host Computer(486 or

Pentium)

Inclinometers

PC

bu

s

Analogue toDigital

Converter

RadioEthernet

DGPS

RS232

Aerial

Metal detector

3-AxisController(Axes 1-3)

D

D

D

D

D

...

DC Motor + Encoder

3-AxisController

(Axes 21-23)

Driver

Motors 2 and 3

DifferentialsystemBody

attachmentFoot

Joint 3

Joint 2

Joint 1 Motor 1

mobile robots . Legged locomotion provides someimportant advantages for moving in natural terrainand seems to be a good solution to carry minesensors over infested fields in an efficient way.Some preliminary work has been developed instudying the potential possibilities of using walkingrobots for demining. This paper addresses thedevelopment of a walking robot endowed with amanipulator able to scan areas with a metal detector.The potential advantages of walking robots arepresented and a new system for land mine detectionis presented . This paper introduces the main systemand some details of the walking robot andmanipulator configurations are provided. Hardwareand software architecture are sketched.The system has to be completed with the requiredtools to form databases of the potential alarms as wellas providing images and graphs to the operator. Theincorporation of new sensors, detectors, and softwarefor signature analysis will be provided in a secondproject step.

7. References

[1] Baudoin, Y., Acheroy, M., Piette, M. andSalmon, J.P. “Humanitarian Demining andRobotics”, Mine Action Information CenterJournal . Vol. 3, No. 2, Summer 1999.

[2] Galvez, J.A., Estremera, J. and Gonzalez deSantos, P., "SILO4-a versatile quadruped robotfor research in force distribution," in Proc. 3rd

International Conference on Climbing andWalking Robots and the Support Technologiesfor Mobile Machines, Professional EngineeringPublisher, U.K., pp. 371-383, 2000.

[3] Gonzalez de Santos, P. and Jimenez, M.A.,"Generation of discontinuous gaits forquadruped walking machines," Journal ofRobotics Systems, Vol. 12, No. 9, pp. 599-611,1995.

[4] Gonzalez de Santos, P., Armada, M.A. andJimenez, M.A, "Ship building with ROWER,"IEEE Robotics and Automation Magazine, Vol.7, No. 4, pp. 35-43, 2000.

[5] Grieco, J.C., Prieto, M., Armada, M.A. andGonzalez de Santos, P., "A six-legged climbingrobot for high payloads," in Proc. IEEEInternational Conference on ControlApplications, Trieste, Italy, pp. 446-450, 1998.

[6] Habumuremyi, J.C., "Rational designing of anelectropneumatic robot for mine detection",Proceedings of the 1st International Conferenceon Climbing and Walking Robots, pp. 267-273,Brussels, Belgium, 1998.

[7] Hirose, S. and Kato, K., "Quadruped walkingrobot to perform mine detection and removaltask", Proceedings of the 1st InternationalConference on Climbing and Walking Robots,pp. 261-266, Brussels, Belgium, 1998.

[8] Klein, C.A.; Kittivatcharapong, S.: “OptimalForce Distribution for the Legs of a WalkingMachine with Friction Cone Constraints,” IEEETransactions on Robotics and Automation , Vol.6, No. 1, pp. 73-85, 1990.

[9] Marques, L., Rachkov, M., and Almeida, A.T.,"Control system of a demining robot",Proceedings of the 10th MediterraneanConference on Control and Automation, Lisbon,Portugal, July 9-12, 2002.

[10] Nonami, K., Huang, Q.J., Komizo, D., Shimoi,N., and Uchida, H., "Humanitarian minedetection six-legged walking robot",Proceedings of the 3rd International Conferenceon Climbing and Walking Robots, pp. 861-868,Madrid, Spain, 2000.

[11] Song, S.M. and Waldron, K.J, MACHINESTHAT WALK: The Adaptive SuspensionVehicle. The MIT Press Series in AI, 1989.

Figure 7. Software architecture

Gait Generationand

Stability Monitor

CommunicationsWalking robot

ParametersTrajectories

ManipulatorParametersTrajectories Motion Processes

Terrain AdaptationAltitude ControlAttitude Control

Trajectory ControlRobot Kinematics

(WALK library)

Leg Kinematics(KIN library)

Leg Control(LEG library)

Axis Control(LINK library)

Hardware Interface(ROB and LM libraries)

Equipment andSensor DataAcquisition

Manipulator KinematicsTrajectory Generation

(KIN library)

Axis Control(LINK library)

Hardware Interface(ROB and LM libraries)

INTEREST OF WALKING ROBOTS IN HUMANITARIAN DEMINING PROJECT

J-C. Habumuremyi, I. Doroftei, Y. Baudoin

Department of Applied MechanicsRoyal Military Academy

30 Renaissance Av. 1000 Brussels, Belgium

Abstract

Three ways (which are more used) for obtaininglocomotion of a vehicle are wheels, tracks and legs.Wheels have the advantages of engineeringsimplicity, low friction on a smooth surface and canmove at high speed. Tracks are a way of extendingthe use of wheels to soft and rough ground. The mainweakness of the two types of locomotion is their poorperformance when faced with a vertical step or adiscontinuous surface. This weakness can beovercome by using walking machines. In this paper,we describe the necessity of using walking robot incertain situations of the humanitarian deminingwhere wheels and tracks can not operate correctlyand we also describe the legged prototypes robots wedeveloped for carrying the detection-sensors to betested on minefields, namely AMRU4 and AMRU5.

Keywords: humanitarian demining, walking robot,legged Robot, Control, Mechanical design

1. Introduction

Even if hostilities ended, mines continue to makecasualty. In accordance with statistics, each yearmore than 26000 people are killed, wounded,disabled by mines. Many mines usually used didn’thave auto-destruction or auto-neutralizationmechanism. From that moment, those explosiveobjects are active weapons even after the end ofmilitary conflicts. About 150 millions of mine arescattered1 and each year 2 to 2.5 millions are placed.In the same time, the United Nations Organizationlocalizes and destroys about 80000 to 100000 mineswith intensive efforts. Against the gravity of thisproblem, one must:1- ban the manufacturing, sale and use of mines2- develop new technologies in mine clearanceThe first point, which is political, is on the good way.Many governments, Non-GovernmentalOrganizations, International Comity of the Red Crossmake efforts in mind to ban mines. The second pointis our concern. We belong to a project, initiated byThe Belgium Defense Ministry. And the task of our

1 Numbers mentioned came from the data base ondemining (1997) of the humanitarian affairs department ofUnited Nations

Robotic Lab is to analyze and to develop low costrobotic demining procedure [1]. For that, we designand build different kind of robots (wheeled, tracked,legged) because one type can’t solve this complexproblem. Wheels have advantages of engineeringsimplicity, low friction on a smooth surface and canmove at high speed. Tracks are a way of extendingthe use of wheels to soft and rough ground by layingdown a track for wheels to run on. But wheels andtrack have a main weakness. They have poorperformance in an unstructured environment whenfaced with a vertical step or a discontinuous surface(which is the structure of the most fields infestedwith mines). Legged robot is a candidate forovercoming this weakness. It has also otheradvantages, as it will be described in this paper. Wealso describe the legged prototype robots wedeveloped for carrying the detection-sensors to betested on minefields, namely AMRU4 and AMRU5.Because big problems of mines are in poor countryand money gave to Non-Governmental Organizationsare limited; we must take this aspect into accountwhen we design such robots.

2. Why to use legged robot in humanitariandemining?

We can justify the use of legged robot inhumanitarian demining by the known reasons [2]above:- After mine clearance, the ground must be useful

for agriculture. The use of wheels and tracks willdamage strongly the ground. Legs do lessdamage to the ground than wheels and tracks.

- The contact with the ground is discrete. We canin this case avoid putting the leg on the mine,avoiding obstacle, and following irregularities ofthe ground by choosing where to put its feet.

- The body height and the posture of the leggedrobot can be changed in according with theterrain that it explores.

- On soft ground a wheel or track will climb out ofa rut of its own making. This lead to wastespower. In an extreme situation, the wheel or thetrack may just dig itself deeper until the vehiclestops. On the other hand, legged robot usescompaction resistance as part of its source ofthrust.

Those reasons have driven us to investigate on the

use of legged robot within the scope of thehumanitarian demining. Below we describe somerobots we built.

3. Sliding Robot AMRU4

3.1. Mechanical structure of the robot

3.1.1. Description

Figure 1: Sliding Robot AMRU4

The AMRU4 Robot (Figure 1) has three functionalparts used:- In the mobility of the robot by means of two

double-acting rodeless cylinder- In orienting the robot in a certain direction by

means of a pneumatic rotary drive- In scanning the ground surface. This part is

placed above the two parts described above andhas an arm where a metal detector or GroundPenetrating Radar (GPR) is suspended. We usedtwo double-acting cylinder and two incrementallinear scales for positioning or for having theposition of the metal detector or the GPR.

3.1.2. Locomotion

Locomotion requires a contact between the mobileand the ground which resist to forces normal to thesurfaces, thus providing support and adhesion. Thiscontact must be powerable so that a propulsion forcecan be exerted across it. The most basic mechanismsfitting this requirement are the slide, the lever and thewheel. The slide mechanism has been chosen for ourrobot. The advantage of such robot is that slidingframe control system is less complex and thecorresponding propulsion system is more robustcompared to lever mechanism. Combining the slidemechanism and the pneumatic actuation, we avoidsome of main origins of energy losses in leggedlocomotion such as:- Power wasted in supporting the body against

gravity (see below on the particularity of thepneumatic circuit of the leg)

- Geometric work because we use a kind oforthogonal coordinate mechanism which

improves the Gravitational Decoupled Actuation

3.1.3. Displacement principle of the robot

A simplified form of the displacement principle isshown in Figure 2.The innerframe, with fourvertically double-acting rodless cylinder used as legs,slides longitudinally on the bed of the outerframe bymean of two double-acting rodless cylinder. Theouterframe has also four legs. To walk, the machinestands alternately on the outer and inner sets. That isan alternating tetrapod gait. The step length is 20 cmand the leg cylinder stroke is at the moment 8 cm.

Figure 2:Displacement principle of the robot

3.1.4. Change of the direction

We use a steering mechanism separate from gaitgeneration to provide the change of the direction.Figure 3 shows on a plane view, how this has beendone. Each set of four legs is attached to a frame,which moves as a whole. The two frames can rotaterelative to each other by mean of a rotary drive abouta vertical axis. To turn the inner frame is firstpositioned at a center of the outer frame using aproximity sensor. Then, the robot stands on one set oflegs and rotates the frame bearing the raised legs.

Figure 3: How to change the direction of the robot

3.2. Pneumatic part

We only describe the pneumatic circuit of legs, otherscircuits are similar or classical. The circuit of eachleg, as shown in Figure 4, has 2 flow control valvesadjustable in one direction, 2 piloted non-returnvalves, 2 3/2-way valves (3 connections and 2switching positions) actuated by a signal applied on asolenoid. The 3/2 valves reset via a pneumatic springonce this signal has been switched off. If valves A1and B1 are in position shown in figure 4, the air inthe two chambers of the piston will be blocked (wemaintain the robot in position even if the compressordoesn't work and we also didn't waste power insupporting the body of the robot). If a signal is forexample applied on the valve A1, it will switch its

position. The piloted air coming from thecompressor, via A1, will maintain the valve B3 open.The air in the chamber 1 of the piston will beexhaust to the atmosphere meanwhile the compressair will enter in the chamber 2. The piston will go upif the signal is applied on the valve B1.

Figure 4: Pneumatic circuit of one leg

3.3. Electronic Part

We have used a Motorola microcontroller MC68332(Tattletale 8) to control the whole robot. The mainspecifications of this �C are: 256K RAM, 256KEEPROM, 12 bit A/D converter, less than 25 digitalI/O lines, 14 TPU (Time Processing Unit) lines, Cprogramming language. The �C has 25 digital I/Olines that can be configured to work either as input oras output. 25 lines are not sufficient in ourapplication, which requires about 60 lines (25 outputsand 35 inputs). That's why we have designed a boardto extend them. The principle is the use of a set oftristate buffers (74HCT541) as digital input device.They are attached to the data bus and enabled byaddress decoding to respond to a specific addressfrom the �C, and clocked in by the �C write pulse.We have also made interface circuits betweenpneumatic components and sensors.

3.4. Positioning and attitude control

The choice of pneumatic are of interest forhumanitarian demining application because they areeconomical (large power output obtained at relativelylow cost), easy to assemble (simple nylon tube andpush-in fittings can be used at the usual pressure of 7to 14 bar), clean (no oil to leak), less sensitive toleakage, can be driven at high speed,Unfortunately, since air is an elastic medium, apneumatic actuator cannot be treated as rigid. Thisimplies that servo control of position is difficult,particularly, if the load fluctuates, as the actuator willbehave as a spring. The solution is the use ofproportional valves, which is more expensive. Thissolution is not acceptable because the robot will costtoo much. But we have found a way of modelingpneumatic actuator and a cheap control strategy [3]for on-off valve based upon these models.The strategy control is as follow: If we would like to

position the piston at a range dx for example (Figure

5), knowing initial values kx , kx� , � �kp1 , and

� �kp2 , we can obtain x� such that if we switch off

valves at range xxd �� , we'll reach closely dx .

x� is computerized in this way:- The microcontroller measure )1( �kx by using

its TPU lines and the optical encoder- The �C calculates )(),(),1( 21 kpkpkx ��� � . Those

terms are expressed in the discretization form ofequations (equations of the dynamic model ofpistons) below:

� �1

)sgn(0

)(

)(2

01000

0)(0

00)(

410

10

2

1

20

203

20

10

101

10

2

1

�������

�

�

�������

�

�

�

�

����

�

�

����

�

�

�������

�

�

�������

�

�

��

�

����

�

�

����

�

�

xF

kkVRT

kkVRT

xxpp

Mb

MA

MA

VApkk

VRT

VApkk

VRT

xxpp

dtd

c

s

s

s

�

�

��

�

�

��

��

Where 321 ),(),( kkkkk and )(3 kk are parameterswhich will depend of the operating point.- x� is calculate with the formula (we consider

the behavior of the air as isothermal):

)2(1

1

1

1

2

x

xxc

pp

cx �

���

����

� ��

- The microcontroller tests if it reaches the rangexxd �� and switch off valves when the

condition is true.

Figure 5: Variation of the range when valves are closed

We have used a FAS-A inclinometer (MicroStrainProducts) to control the attitude of the robot.

4. Hexagonal legged Robot AMRU5

4.1. Mechanical Design

4.1.1. Characteristics and structure of theproposed leg

We have made a study of various types of legarchitecture [4] from different robots such as Silexrobot (developed at Free University of Belgium),Lauron robot (developed by FZI at KarlsruheUniversity in Germany), etc. These studies lead us to

choose leg based on 2D-pantograph mechanism forthe robot we intend to build (Figure 6 and 7). Wehave chosen that structure of leg by consideringadvantages and disadvantages of those legs. Itsadvantages are as follow: It generates an exactstraight-line on the two axis of the Cartesiancoordinate system; this means that for walking on asmooth terrain no more than two motors are actuatedat the same time. It has good energy efficiency (lessenergy consumption due to the completely decoupledfoot motions and exact straight-line motioncharacteristic), good rigidity, simpler controlalgorithm (due to the completely decoupled motionand linear relationship between input and outputmotions). As disadvantages we can mention thelinear actuating systems are more difficult to designand to protect from the environment and may havemechanical reliability problem.

Figure 6: Leg based onpantograph mechanism

Figure 7: Prototype designof the leg

Figure 8 and 9 show the whole robot respectively asdesign and as built.

Figure 8: Design of theAMRU5 Robot

Figure 9: General view ofthe robot AMRU5

4.2. Control of the robot AMRU5 (underdevelopment)

Control of a walking robot is very important. Withouta good control, we can’t reach advantages of walkingmachines even if the mechanical structure was welldesigned. That’s why, we’ll linger on that aspect.By considering the environment and elements, whichinteract with the control algorithm of a walkingrobot, we expect that such algorithm provide abovefunctions:- stabilize the robot- know geometric configurations of legs to be use

in feedback or in another strategy of control- coordinate movement of different legs

- To allow the robot to walk in different kind ofterrain (hard, soft)

- To run over or to avoid obstaclesThose functions require sensors. In our case, we haveused inclinometers (for the stability, attitude andlatitude control), optical encoders and switches (forthe geometric configurations of legs) and finallyforce sensors (for the robot to walk in hard and softterrain, also for obstacle avoidance).

4.2.1. Classical modeling or cognitive modeling?

Advanced classical control schemes for robotmanipulators often involve computation of the jointactuator torque from its motion by using a dynamicmodel (physical modelling). The dynamic model isexpected to properly predict the joint torque fordifferent trajectories and load conditions, in real-time. If this model has been found, the well-knowntechniques of approximation, linearization,validation, and stability analysis are applied to obtainan appropriate control. Whenever, it seems difficultto have such a model due to the problem for havingaccurate internal parameters of the robot (distancebetween joints,) and it is not easy to accurately modelsome complex phenomena (nonlinearities) such asbackslash, friction, flexibility, … The problembecomes more complex for a walking robot used inunstructured environment and specially inhumanitarian demining application. Such robot has agreat number of degree of freedom and requireschanging internal parameters depending on theenvironment that it explore. In this case, cognitivemodelling such as Fuzzy Control seems to bereasonable if experience has shown that theunderlying control task is feasible by human experts.This explains the reason we have chosen cognitivemodelling to design the controller of a six-leggedwaking robot instead of classical modelling.

4.2.2. Adaptive Neuro-Fuzzy Controller

Fuzzy Logic Controller (FLC) lets you describedesired system behaviour with simple “if then”relations. You can, with FLC, use all availableengineering know-how to optimise the systemperformance directly. But the problem with FLC ishow do derive rules from data sets. On the otherhand, Artificial Neural Networks (ANN) are able toapproximate or to solve certain tasks by learningfrom examples. When data sets contain knowledgeabout the system to be designed, an ANN promises asolution because it can train itself from the data sets.But the problem with ANN is that it remains a “blackbox”. A clever combination of the two technologies(FLC and ANN) produces the Neuro-Fuzzy. Neuro-Fuzzy [5] combine the explicit knowledgerepresentation of fuzzy logic with the learning powerof neural nets.

4.2.3. Implementation of a Neuro-Fuzzy controller

To control the robot AMRU5, we have chose a well-known hierarchical control (Figure 10) [6]. The highlevel sensory input produces a motivation to move byproviding a high level command which can be in ourcase “found a mine”, “avoid a mine”. The high levelcommand is then converted into linear and angularvelocity of the centred frame of the robot (body routelevel). Then kinematics planning level produce jointspace vectors for lower-level actuators. Those vectorsare after used in the control of actuators by includinglow-level sensory input. At the moment, we are moreconcerned with kinematics planning specially on legcontrol. For that, we have defined at each joint twoANFIS (Adaptive Neuro-Fuzzy Inference System)controller. One will be used in support phase and theother in transfer phase.

Figure 10: Hierarchical control

Below we present an original way of implementingNeuro-Fuzzy approach on each leg [7].

4.2.4. Neuro-Fuzzy approach on an existingFuzzy controllerThree steps below, was applied to design an adaptiveNeuro-Fuzzy:1- Design of a Mamdani fuzzy controller which

seems easy to fix2- Copy of the existing approximate controller to

make it Sugeno

3- Realisation of adaptive Neuro-Fuzzy controller

4.2.4.1. Design of a Fuzzy PD controller

As an example, we have chosen a Mamdani FuzzyPD controller but it can be any other controller that iseasy to fix. The Fuzzy PD controller derives from thestandard (classical) PD. The output of theconventional analogue PD controller in the frequencys-domain is given by:

Where cpK and c

dK are respectively the proportional

and derivative gains and )(sE is the tracking errorsignal. By applying the bilinear transformation

112

�

��

zz

Ts (where T is the sampling period) and then

taking the inverse z-transform, we have:

��

���

� �����

���

� ��

TneneK

TneneKnu dpPD

)1()()1()()( (3)

Where cpp KK � ,

TK

Kcd

d2

� and

Tnunu

u PDPDPD

)1()( ���� (4)

We can then rewrite (4) as:)()1()( nuTnunu PDPDPD ����� (5)

If we replace the term )(nuT PD� by a fuzzy controlaction gain, we finally obtain

)()1()( nuKnunu PDuPDPD ����� (6)

Where uK is a fuzzy control gain. Figure 11 showthe Fuzzy PD controller.

Figure 11: Fuzzy PD

We can also rewrite (3) as:)()()( nrKndKnu dpPD ���

Where T

nenend )1()()( ��� and

Tnenenr )1()()( ��

�

4.2.4.2. Copy of the controller above

After the design of the controller, we transform it in afirst order Sugeno fuzzy. A first order Sugeno fuzzyis not easy to design because rules have this form:If )(1 ndKinput p� and )(2 nrKinput d� then

� � � � � � � �isnrKiqndKipiu dp ���� )()(

Where � � � �iqip , and � �is (i between 0 and numberof membership function power to the number ofinput) are parameters to fix in addition of pK ,

dK and

uK . But optimisation techniques can be applied on aSugeno Fuzzy. To copy the controller, we collect inreal-time datas )(ndK p , )(nrKd and u� whenrunning the existing controller (Figure 12). After, wetrain off-line an ANFIS controller for to fixparameters of membership function and of Sugenofuzzy controller. This is done using well-known

)()()( sEsKKsu cd

cpPD ��

optimisation techniques in Neural Network such asbackpropagation or LSE methods.

Figure 12: Copy of an existing controller

4.2.4.3. Realisation of the Adaptive Neuro-Fuzzycontroller

Figure 13: Adaptive Neuro-Fuzzy controller

With the off-line training, we have just anapproximate controller. On-line training allow theadjustment of parameters to enhance its performance.In our case, parameters of the controller must changedepending on the environment that the robotexplores. We did that by duplicating the inversemodel controller. Figure 13, shows two ANFIScontrollers, but in fact, they are the same one, theduplication reflects the fact that two cycles arerequired.

4. Conclusion

AMRU4 Robot is a low cost and lightweight mobileprototype-sliding robot. It displaces step-by-step a3D scanner along parallel corridor defined on aregular minefield. On the arm of this scanner can beplaced a metal detector, a ground penetrating radar(GPR) or any sensor used in demining. This robothas been tested on a training minefield. This testshown that the concept of this robot is very good butwe must arrange various elements in a proper way tomeet requirements and constraints of a reallydemining robot e.g. 8 cm of leg’s stroke is notsufficient for to walk on a irregular terrain.AMRU5 Robot has been built after a survey ofexisting legged robot. The choice of legs based onpantograph mechanism solved many problems e.g.rigidity of the robot, control less complex, And wealso think that cognitive modelling can be a goodway for to built an intelligent machine. Resultsobtained on a prototype leg when testing the adaptiveNeuro-Fuzzy had given satisfaction. We are stillworking on the control of the AMRU5 robot. Thefuture will reveal if with such control, we’ll be able

to have a robot we really move in unstructuredenvironment. After that an arm can be placed on therobot. On this arm, usually sensors used in deminingcan be fixed.

5. References

[1] Y. Baudoin and E.Colon (1998), “HumanitarianDemining and Robotics”, IEEE on ControlApplication, Sept 1-4,1998

[2] P. Alexandre (1997), “Le contrôle Hiérarchiséd’un Robot Marcheur Hexapode”, Phd Thesis,pp. 1-10

[3] J-C. Habumuremyi and Y. Baudoin (2000),“Control of an Electropneumatic Sliding Robot”,Clawar2000, pp. 507-515

[4] J-C. Habumuremyi and I. Doroftei (2001),“Mechanical Design and MANFIS Control of a Legfor a New Demining Walking Robot”, Clawar2001,pp. 457-464

[5] D. Nauck, F. Klawonn and R. Kruse (1993),“Combining Neural Networks and FuzzyControllers”, FLAI’93, Linz, Austria, Jun. 28-Jul. 2

[6] M.J. Randall (1999), “Stable Adaptive NeuralControl Of Systems With Closed Kinematic ChainsApplied to Biologically-Inspired Walking Robots”,PhD Thesis

[7] J-C. Habumuremyi, I. Doroftei and E. Colon(2002), “Adaptive Neuro-Fuzzy Implementation ForLeg Control of a Walking Machine”, Clawar2002

MINE CLEARANCE ROBOTS

Štefan Havlík

Institute of Informatics, Slovak Academy of Sciences,Severná 5, 974 01 Banská Bystrica, Slovakia,

E-mail: havlik@savbb.sk

AbstractActual task of detection and neutralization of landmines abandoned on post battlefields is a big chal-lenge for R&D in many fields of robotics. This paperbriefly surveys problems, analyses some importanttasks could be performed by robotic systems. Someconcepts of solving such robotic systems are out-lined.

Keywords: Demining, Mobile Robotics, RemoteControl, Sensing,

1. Introduction

Remarkable evolution of land mine technology in lastyears resulted in development of new plastic mines,as well as those containing a minimum amount ofmetallic parts. Mines are cheaper, more reliable,more durable and harder to detect and dismantle.Remote deployment systems (such as from helicop-ters) have made it possible to deliver thousands ofmines to a broad territory within a short time. Newmaterials and laying mines in this way makes itimpossible to detect them and to record where theyland, so recovering them is getting more difficult.

Classic methods for detection and removing mines,used at present, are dangerous, too costly and consid-ering the number of abandoned mines, are very slow.Within technologies for cleaning large mine pollutedareas most frequently used are mechanical systems.Usually a pressure is acting on the ground by rollerspushed ahead of a tank or rotary flails with hammersbeating the ground, or, the mines are dug out andpressed by a rake. Main drawback of these purelymechanical techniques is that they should mechani-cally effect on large areas, frequently, without anyoccurrence of mines. More, no such system cansatisfy desired 100% reliability of humanitarian minecleaning and frequently manual verification of yetcleaned area is required. For these reasons there is anurgent need to develop fast, safe and efficient de-mining methods. It should be said that the mainproblem of demining lies and will be solved if minesare reliably detected and precisely localized. Then theneutralization procedure is directly addressed to thisplace of mine occurrence. Both steps could be madeby automatic ways. This is a big challenge for roboticresearch.

It is generally required that such a robotic system fordemining should work autonomously i.e. it is able tomaneuver in various terrain, it has to detect mines, orany other explosives and then to perform theirdestruction.Research includes development of reliable detectionsystems, new – task oriented design of robots, as wellas sophisticated control methods. Robotic systemshave to be equipped by tools for performing specialdemining tasks and technologies.Much research work has been yet done in the domainof detection and localization of mines. Beside knownmethods new sophisticated sensing principles able todetect and recognize mines as hidden objects areunder development [1,2,9]. This is most crucial taskin the whole process. Because of automatic deminingprocess is based on using special robotic vehicles /agents further research is oriented to the developmentmobile agents able to operate in / above the danger-ous and partially unknown terrain as porters ofdetection systems and other demining tools. Someconcepts and problems of robotics related to terres-trial demining are discussed below.

2. Demining Techniques And Systems

In general, the mine cleaning procedure consists oftwo main tasks:

- Detection and localization of land mines.- Neutralization i.e. removing or destruction of

mines on place.Both these tasks are directly related to the problemand solving third important task:

- Preparing infected terrain for reliable detection aswell as for neutralization procedures, i.e. re-moving vegetation and any obstacles that couldprevent detection or safe neutralization.

Considering most general robotic mine clearancetechnology an advanced system consists of followingparts.

Mobility system represented by remotely controlled /autonomous / semiautonomous mobile (robotic)vehicle as general porter of platforms and othersystems for performing three above principal tasks.In principle, there are possibilities as follows:§ Free flying vehicles with suspended platform i.e.

airborne mission mainly for searching especiallylarge area

§ Ground vehicles (wheeled / belt or walking /legged machines)§ Cable suspended platforms moving over the

dangerous terrain

Multi-sensorial system for detection / recognition ofmines. [1,2,9] As to sensing principles automaticdetection techniques should satisfy reliable detection/ recognition of mines and to mark them into maps(assign them coordinates) as targets. In principle, thesensory systems can be situated on a special platformof a mobile vehicle performing scanning dangerousterrain or on the end of a robotic arm. The final goalis localize all objects have been detected and recog-nized as mines and to write their coordinates intomaps.Despite very strong and expensive research the effortto develop a sophisticated low-cost sensing systemsdue to large variety of mines and terrain no singleprinciple can be used but fusion of several detectionsshould be applied for reliable recognition. Withinsensing principles and available technologies (metaldetection, infrared imaging, ultrasound, etc.), otheractive / passive principles are under elaboration.There are for instance: Ground Penetrating Radar(GPR) / Mm- wave radar / ultra-wideband radar, X-ray spectroscopy, active acoustic and seismic, mag-netic field sensing, neutron activation analysis,charged particle detection / IMS (Ion MobilitySpectroscopy), nuclear quadrupole resonance,chemical, biosensors, bacteria. As obvious, someprinciples are more suited for searching large areas todetect the existence of minefields (infrared, chemical,bacteria) and the others should enable precise locali-zations of particular targets.Tools for neutralization / destruction of mines.[2,4,10]

Beside mechanical systems as for instance: rollers,ploughs, flails, rakes, hammers, …etc, other princi-ples that activate explosion of mines can be used.There are: explosive hoses, fuel air mixture, directedenergy systems, laser, microwave sources or sniperrifle. For the mine removal tasks there are end-effectors in forms of double shovels, diggers, etc.Input data for these systems are positions / coordi-nates of mine targets.Tools for removing obstacles / vegetation andpreparing terrain. Mines after some years of de-ployment are covered by sand (in desert conditions),ground, vegetation, masking means, …etc. Forremoving these obstacles different end of arm tele-operated tools with sensory feedback should bedeveloped. There are: sand suckers, cutters, shovels,special grippers, diggers and probes, etc.

3. Some Concepts of Robotic Tools

Considering large polluted areas and drawbacks ofactual demining technologies main contributions

from using robotic technology in mine cleaning areexpected especially in following topics:

− Searching large areas and localization of mines andany explosives (UXO) by fast and reliable way.

− Reliable neutralization/destruction of mines with-out the need of personal assistance to be inside orclose to the dangerous place.

3.1. Global concept “ANGEL” [1]

A most general concept considers activity of twomissions: aerial and ground, having a commonoperation / information center. Main function of thisoperation center is to collect information, planningactivities and evaluation of actual situations as wellas controlling agents for detection and neutralization.The system operates with GPS data over digital GISmaps.

The agents for performing these tasks can be, inprinciple, aerial or ground vehicles that satisfydesired mobility features and are provided by ade-quate technology equipment, depicted in Fig.1.

Fig.1. Global concept of demining

Flying vehicle. Aerial searching is especially suitedfor first scan of large areas. Unmanned flying vehicle– small helicopter for this purpose is equipped by aspecial platform with several detection systems. Thehelicopter performs scanning motions over the terrainand as soon as any suspicion on mine (UXO) willarise coordinates of this place are saved into opera-tion map. More precise localization of particularmines is doing by ground detection vehicle in nextsearching.

Ground vehicle for detection. The vehicle automati-cally moves to address from digital map where anexplosive was observed / detected by aerial search-ing. Its task is localizing exact position of mines, and/ or to mark them by a visible color. The vehicle asporter of multi-sensorial system should exhibit verygood maneuvering and control capability in variousterrains as well as autonomy features that allowcollision / mine avoidance, automatic stop in cases ofdetected mine, remote vision, etc. From the point ofview mechanical performance there are some severalspecific requirements that such a vehicle shouldsatisfy (maximal pressure on the ground, velocityrelated to speed of detection systems, noise and

Target

temperature limitation, reliable power and communi-cation systems, self recovery capabilities, etc).

Neutralization – mine destruction vehicle. Thisvehicle with similar maneuvering and control capa-bilities as the above has to approach to the position ofa detected mine and to neutralize it by activation orremoving procedure. For this reason it has to beprotected against explosions of mines not onlyantipersonnel but anti-tanks too.

3.2. Platform for searching dangerous terrain

For searching dangerous terrain and relatively largeoperation space the concept of the cable suspendedrobotic platform was designed [6,8].

In principle the system, as schematically depicted inFig.2, consists of three cable winches fixed onmobile columns. The ends of cables from particularwinches are connected on the platform moving abovethe working place. Each winch mechanism isequipped by the cable length measuring sensor andthe position / velocity control. Thus for such aparallel mechanical system any actual position of themoving platform determine three distances i.e.measured lengths of cables between the platform andend pulleys of winch mechanisms. The centralcontrol system performs transformations and coordi-nated motion control of the platform with respect to aworld reference frame defined on place.

Fig.2. Concept of searching dangerous terrain

This concept of scanning dangerous terrain by thesuspended platform with detectors exhibits someadvantageous features as follows:- Large workspace of operation, reconfigurable

according to actual terrain conditions- Low weight and simple transport- Fast and simple installation on place- Operation / control in Cartesian coordinates de-

fined directly on place.

Four main problems have been solved for this sys-tem. There are:- Kinematic and force analysis (motion and force

transformation i.e. functions that relate actual mo-tion and load values expressed in world referencecoordinates and internal representation of controlparameters i.e. cable length and internal forces).

- Coordinated motion control in world coordinates.

- Dynamic analysis and control.- Calibration, i.e. to actualization of parameters in

relations for motion and force transformationsaccording to real configuration of the system andits spatial geometry.

Let us briefly analyze the calibration procedure whatrepresents an important SW tool in scanning thesearched area as well as mapping targets in worldcoordinates. In practice, when all three winchmechanisms on vehicles have been installed on placespatial coordinates of end pulleys i.e. fixation pointsA,B,C should be determined. Principal requirementis to perform this calibration without any additionalequipment.

The procedure, as proposed [6], consists of foursteps:

a) Let us stake out three points M, N, and P on theground x-y plane. These points create a trianglewith known geometry. Although, in principle, thistriangle could be chosen quite arbitrarily moreadvantageous will be to construct it equilateral, asdepicted in Fig.3.

b) Using individual command of particular servos weperform positioning of the platform sequentiallyinto points M, N and P. Denote by symbols: qAM,qBM, qCM, qAN, qBN,, qCN and qAP, qBP, qCP allmeasured lengths of cables that correspond toparticular positions according to Fig.3.

c) Solving three tetrahedrons MNPA; MNPB andMNPC, the coordinates x,y,z of A,B,C points arecalculated.

d) Actualize the transformation matrix T0A

that

relates actual configuration of fixation pointsA,B,C with respect to a given ground referencesystem O(x,y,z).

Os

ss

x

yz

N

P

M

A

B

C

qAM

qAP

qAN

qCM

qCNq CP

qB N

qBP

Fig.3 Geometry for calibration

3.3. Vehicle for searching dangerous terrain andtarget localization.

Main function of this vehicle is a remotely / pro-grammable controlled general porter of severaldetection systems. It exhibits excellent mobility andmaneuvering capabilities in various terrains. As thefunctional equipment of this mobile robot there areadditional tools with features as follows:

O

Scanning motion

B

C

A

§ The robotic arm with end platform equipped bydetectors§ The long reach 6 d.o.f. robotic arm with the set of

exchangeable tools (grippers, cutters of vegeta-tions, detectors, etc.).§ The remote vision system with the camera fixed on

pan-tilt unit.§ Marking system for detected targets.§ Sensory systems for collision avoidance and

navigation.§ Control and communication system.

Remark: It should be noted that destination of thevehicle to work in dangerous environments requiressome specific systems and protection equipment. Acrucial importance plays self-recovery system.

Fig.4. Detection vehicle with sensory platform androbot hand

3.4. Neutralization / destruction vehicle.

The neutralization vehicle is the general porter ofvarious destruction systems (flails, guns, laser gun,explosives, etc.). The vehicle with similar maneu-vering capabilities should be more robust and beprotected against of the pressure waves duringexplosions of mines. A principal configuration isshown in Fig. 5.

Fig. 5. Vehicle with flailing destruction system

The new realization of this vehicle named “Diana’with flailing destruction tool can be seen in Fig.6.[11].

Fig.6. The “DIANA” flailing vehicle for demining

Flailing mechanisms

Flailing mechanisms is a special equipment fordestruction mines i.e. activation of explosions bymechanical way. Hammers on the ends of chainsrotate with the shaft and hit to terrain by. Impact

force of hammers results in explosions of APM andATM. The system is fixed on the manipulator flangein front of the vehicle. During operation the endflange with flailing systems are controlled accordingto the terrain by adaptive way.

Robot handThe on-board robot arm performs some specific tasksespecially in situations as follows:− Targets are not exactly localized and additional

searching / detection by robot hand held detectorsshould be made.

− Obstacles (vegetation, stones, etc.) that preventlocalization / neutralization should be removed.

− Using special tools for neutralization.

The robot hand with the reach 3 m has payloadcapacity about 20 kg. It could be controlled in Carte-sian hand and vehicle reference coordinates related tocamera systems. Concept of this robot hand inworking and transport positions is in Fig. 7.

Fig. 7. Robot hand on the vehicle

3.5. Vehicle operation system

As discussed above, searching and demining proce-dures made by robotic vehicles – mobile robots thatexhibit some level of autonomy. This fact naturallyrequires some unified approach to navigation andcontrol of particular vehiclesSpecific working conditions for vehicles and robotictools and security reason require that the controlsystem to work in two independent modes:§ Automatic / programmable control mode through

communication with operation center. This modesupposes normal operation of all systems as in-cluded scheme in Fig.8. Communication systemfor automatic modes transmit control and sensorydata: way-points / trajectory, control statements forvehicle and motor, images from camera (remotevision), vehicle and motor states, warning errorsituations, etc.§ Manual control using joystick / control panel /

keyboard that allows maneuvering the vehicleswithout operation center. Manual control will beapplied in cases as follows: removing the vehiclefrom the minefield and recovery of any situationdue to failure of any system (programs, communi-cation, etc.), loading / unloading the vehicles dur-ing transport, testing. This control mode directlyoperates with steering and motor control loops.Communication is limited and corresponds to mainstatements for maneuvering

Fig. 8. Components of the vehicle control system

In general, any demining procedure includes someprincipal control routines. There are:

Task 1. Position and orientation of the vehicle.

Altitude and longitude of the vehicle is directlymeasured by on board GPS unit. The accuracy andresolution of measurements should correspond toaccuracy of digital maps where all targets are re-corded. As to orientation angle (azimuth ϕ) can bedirectly measured by digital compass. Then, threevariables (xV, yV, ϕV) are controlled coordinates ofthe vehicle as can be seen in Fig. 9.

O

V

y

x0

x

V

V

V

N = y0

N ϕ

Fig.9. Position / orientation of the vehicle

Task 2. Maneuvering to a given target. (Direct task)

The vehicle should move to a given target coordi-nates in order to localize its position more precisely,or, to destroy it. For the security reason we statearound the target the security measure ρ and theapproach angle ϕap. These parameters should guar-antee that the first “contact“ of the vehicle with anexpected dangerous target be by a detection system,or, by the destruction system. The approach angle ϕapexpresses direction of movement of vehicle from anactual to a specified vicinity of the expected targetposition. The security measure ρ represents theuncertainty of recording targets into digital map asresult of a limited accuracy of localization duringaerial / terrestrial searching. Considering this uncer-tainty or security measure it is expected that thetarget be situated inside the circle given by coordi-nates in digital map. Then, the searching strategy ofgoal position depends on ϕap and ρ parameters. Sucha situation when the goal position is reached and nextoperation could start is depicted in Fig.10.

O x0

y x

y

x

N

Target

ρV M

M

V

ϑv

N=y0

Vp

p0 V

0 M

M Vp

Mz

M

Fig. 10. Approach to the target

Task 3. Precise localization of target positions.(Inverse task)

The vehicle is in a position and the target is detectedby some of detection systems. The exact position ofthe target should be stated and recorded. Practicallythe vehicle stops in some sensing position andperforms searching dangerous terrain according to agiven searching strategy, which corresponds todetection system just used for searching. (see Fig.11)There are, in principle, two possibilities:− Detectors are on the sensory platform in front ofthe vehicle− Detectors are in the robot hand.The task is then to ascertain positions of targets usingtransformations that relate to particular detectionsystem. As obvious by fusing sensory information itis possible to repeat detection procedure by usingdifferent sensing technologies including visioncamera in the hand. Performing this task the vehicleis then maneuvered to this goal position specified bythree variables xV, yV, ϕap.

O

V

y

y

x

x

x

y

z

V

V

H

H

HH

p

p p

p

0 V

0T HT

V H

T

Fig.11. Precise localization of the target by detectorsin robot hand

Steering control loop

Motor control loop

Tools / arm control systems

OPERATION CENTER

GPS, compass, camera,range finder

Way pointsNavigation to target pointsscanning motions andcollision avoidance strategy

Go: advance / backTurn: left / right, stop,Speed control

Start / stop, Motor patameters

Ground surface tracking,tool control data

wheel sensors,

motor sensors

US/ position sensorshand held camera,...

Communication dataControl statements

Control loopsSensors

Let us briefly describe the procedure for calculationof position of targets in all principal tasks. Consider-ing reference coordinate systems, as specified onfigures the global position of the target detected by asensor can be expressed using transformation

THT H00 pTp ⋅= (1)

where p are positional vectors related to particularreference systems and T is the transformation matrixbetween vectors in these systems

[ ]Tzyx ppp 0,,,=p ,

=

1

VHVH

HV 0pD

T (2)

and DVH is the 3x3 rotation matrix of the H referencesystem into V system and pVH is the positional vectorof the H system with respect to V.Then, considering introduced reference systems it isobviously

VHHTTT 0V0 ⋅= (3)

Because of positions of targets are given by twocoordinates in global – world references particulartransformation matrices can be simplified as follows

−

=

1000

0000

0

0

vVV

vVV

y

x

ϕϕϕϕ

sincos

cossin

0VT (4)

where xv, yv, and ϕv are three measured variables forposition and orientation of the vehicle.Remark. Accurate calculation of target positionrequires that the inclination angle of the vehicle beconsidered. Although the actual inclination of thevehicle can be measured by a two-axis inclinometerfollowing maximal simplicity of sensory equipmentit is reasonable to neglect errors due to inclination.Anyway, when consider maximal allowable inclina-tion angles in terrain this error will be within therange of the accuracy of GPS measurements.

4. Conclusion

The paper presents some conceptual considerationsin designing the control concept of unmanned vehi-cles for detection / localization of mines as well asfor neutralization. It is considered that vehicles areprovided by some degree of autonomy and areprogrammable controlled from the operation centerworking with digital maps and GPS sensory data.Dangerous terrain and avoiding unexpected explo-sions of mines result in applying the specific ap-proach to searching with precise localization oftargets and neutralization. Both operations closelycorrespond to sensory equipment for detection aswell as destruction technology.

References

[1] http://www.eudem.vub.ac.be,www.state.gov/t/pm/rls/rpt/,http://www.hdic.jmu.edu/,http://www.gtd.es/angel_1.asp., www.gichd.ch,http://www.demining.brtrc.com/r_d/, …

[2] Proc. 5 th Int. Symposium on technology and theMine Problem. Monterey, Apr. 21 – 25, 2002,USA

[3] Proc. Int. Workshop on Sustainable HumanitarianDemining, Sept. 29 – Oct. 1, 1997, Zagreb,Croatia

[4] Habib, M.K. Mechanical Mine Clearance Tech-nologies and humanitarian Demining. Applica-bility and Effectiveness.. 5 th. Int. Symp. Tech-nology and mine problem. Monterey, CA, Apr.22-25, 2002

[5] Havlik, S.: Humanitarian demining – a bigchallenge for robotic research. Proc. Int. Work-shop Robotics in Alpe-Adria-Danube RegionRAAD 98, June 26 –28, 1998, Smolenice Castle, Slovakia,

[6] Havlik,S.: A reconfigurable cable crane-robot forlarge workspace operations. Proc. ISIR’93, Nov.4-6, 1993, Tokyo, pp. 529-536

[7] Havlik,S. and Licko, P: Humanitarian demining– the Challenge for Robotic Research, The Jour-nal of Humanitarian Demining, May 1988.

[8] Havlík, Š., Šèepko, P.: Searching dangerousterrain. MINE,99 Proceedings Euroconferenceon: Sensor systems and signal processing tech-niques applied to the detection of mines andunexploded ordnance. October 1-3, 1999, VillaAgape, Firenze, Italy, pp. 72 - 77

[9] Proc. MINE,99 Proceedings Euroconference on:Sensor systems and signal processing techniquesapplied to the detection of mines and unex-ploded ordnance. October 1-3, 1999, VillaAgape, Firenze, Italy,

[10] Hirose, S., Kato, K.: Quadruped walking robotto perform mine detection and removal tasks.Workshop on Anti-personnel Mine Detectionand removal. pp. 30-36

[11] Hontstav, s.r.o., Slovakia, Information materialsabout the new product “DIANA”. 2002.

Rem.: This work is part of the project No. 2/1099/2supported national grant agency VEGA.

ROBOT SWARMS FOR DEMINING – VISION OR REALITY

P. Kopacek

Institute of Handling Devices and Robotics Vienna University of Technology

Favoritenstrasse 9-11, A- 1040 Wien, Austria kopacek@ihrt.tuwien.ac.at

Abstract : Removal of landmines – mostly by hand - is today a very inhuman, dangerous, time consuming and therefore expensive task. On the other hand the number of landmines is increasing worldwide dramatically. Since several years there are several trials to use unintelligent and very heavy robots for demining. In the last decade a new generation of mobile, intelligent and cooperative robots were developed and introduced. One of the reasons for this development was the availability of reasonably cheap sensors and the increasing computer power. This offers the possibility to use robots of this new generation for humanitarian demining. In this paper first ideas were presented and shortly discussed. Keywords: Robots, Multi Agent Systems, Robot Swarms, Demining 1. Introduction Demining can be divided in three steps: - Mine detection: First you have to find ( detect )

and then you must destroy or remove the mines. - Mine digging: To remove the mines from the

ground. - Mine collection: To collect the mines and

transport it to a place where they can be destroyed.

Today used methods for detection are: High-tech methods (Radar, infrared, magnetic tools, touch sensors usually piezo resistive sensors, GPS); Dogs: sniffing the explosive contents of the mines. The commonly used methods for destroying and removal are brutal force methods (include ploughs, rakes, heavy rolls, flails mounted on tanks). Hand-prodding is today the most reliable method of mine removal, but it is a very slow, and extremely dangerous. A person performing this type of clearing can normally only perform this task for twenty minutes before requiring a rest. This method clears one square meter of land in approximately 4 minutes. Mine collection is carried out today mostly manually sometimes with mechanical devices. Therefore (Semi) automatised demining gets more and more important today and in the future. 2. Multi Agent Systems - MAS

A MAS consist of a number of intelligent, co-operative and communicative hardware agents – mobile robots - getting a common task. Because of the intelligence they are able to divide the task in subtasks as long as at least one agent is able to fulfil one subtask. Repeating this procedure yields to the solution of the common task. Newest research goes in the direction of MMAS - Multiple Multi Agent Systems – different MAS are involved for the solution of a complex task. A MAS get a whole task. The host computer divide the whole task in a number of different subtasks as long as a distinct subtask can be carried out by at least one agent. The agents will fulfill their subtasks in a cooperative way until the whole task is solved. Such a global task could be: assemble a car. The agents – mobile, intelligent assembly robots – have to create subtasks (e.g. assembling of wheels, windows, brakes,......) in an optimal way (equal distribution of the workload of the agents) and distribute to the agents. The main hurdles for MAS-research is the complexity of the whole system. This complexity is dramatically increasing by adding new agents. Therefore the interaction, communication, coordination of the tasks between agents, and control are the topics for the development of a Multi Agent System ( MAS ). For heterogeneous robots it is difficult to implement the communication, because each robot has its own kinematic structure, programming language etc.. Furthermore the range of frequencies used for communication and the capability of RF modules is limited. It is necessary to develop standardized communication protocols and methods, that should be one of the work for the next years.

H

A1

A2

An

H : Host computer; A1, A2,.....,An : Agents ________ : full communication ------------ : partial communication

Fig.1a: Communication only between host and agents Fig.1a. shows the present situation of the communication between the host and the agents. For the future the agents should also communicate with the host and also with the other agents which is shown in Fig. 1b and Fig.1c.

H A2

A1

An

Fig.1b: Partial communication between host and agents as well as between the agents

H

An

A1

A2

Fig.1c: Full communication between host and agents as well as between the agents The characteristics of MAS are [4]:

• each agent has incomplete information or capabilities for solving the problem and, thus, has a limited viewpoint

• there is no system global control • data are decentralized • computation is asynchronous

In scientific papers are various approaches and denotations to subdivide control strategies for autonomous mobile robots in different types. There are two fundamental ideas, the functional and the behaviour based approach.

In the functional approach the control system get information about the environment by sensors, constructs an internal model, computes plans to fulfil its tasks, and acts on it. This is the so-called “sense-think-act” cycle. The main disadvantage of this approach is the internal world model, which is ideally a true representation of the real world. However, modelling of the real world is difficult due to problems such as the dynamic nature of the real world, limitations of the sensors and so on. Another problem of the functional approach is that it is extremely brittle. If any module fails, then the whole system will fail. The behaviour based approach avoid this problem by using a parallel structured control system. Here, the overall control task is decomposed into task-achieving behaviours operating in parallel. Each behaviour module implement a complete and functional robot behaviour, rather than one single aspect of an overall control task, and has immediate access to sensors and actuators. The fundamental idea is that task-achieving behaviours operate independently of one another, and that the overall behaviour of the robot emerges through this concurrent operation. Although the system is more flexible and robust, this approach lacks performing complex tasks. A behaviour-based robot responds directly to sensory stimuli, it has no internal state memory and is therefore unable to follow externally specified sequences of actions. 3. Robot Swarms – MAS - for demining Robot swarms improve the capacity of robotic applications in different areas where robots are already used today. Robot swarms are similar to – or a synonym for - ‘Multi Agent Systems – MAS’. These systems are very well known in software engineering – “software agents” - since more than twenty years. In the last years there are more and more works related to “hardware agents” like robots. Applying robots for demining there are two possibilities:

a. using mobile, intelligent multipurpose robots equipped with devices for mine detection, mine removing as well as mine transportation.

b. using three different swarms of single-purpose robots equipped either with detection devices or removing devices or transportation facilities.

Both of these approaches have advantages and limitations. Therefore most of the used control systems are a mix of them. Surely in each combination one of the two approaches is in some measure dominant. These combinations are usually called hybrid strategies. They separate the control

system in two or more communicating but independent components. The lower levels are behaviour based while higher levels follow the functional approach. The goal is to provide quick responses in a dynamic environment while having the ability to plan and perform complex tasks. Our approach is the second one – three different swarms in the minefield. The detection robots scan the field for possible mines. If a metallic part – probably a landmine is detected the removal robot close to this site takes over the removed mine and a transportation robot with free capabilities transports the mine to a collection area out of the minefield. The whole process is fully autonomous. Operators are only needed for monitoring and of course maintenance. To achieve this goal the robots must have a high level of intelligence and must be able to communicate among themselves. Since the power supply of mobile robots is very limited there is also need for docking stations. The host computer in Fig. 2 is necessary to solve the path planning problem in a dynamic environment. Each robot represents for all other robots a dynamic obstacle which has to be avoided. The host computer controls the movements of all robots by means of wireless communication. But soon such a host computer will be obsolete (Fig.1c). Software implemented in the onboard computer of each robot will take over this task. Using different single purpose robots for the different tasks reduces the weight of the robots. Therefore it is much easier to design robots which are lightweight enough not to cause an explosion while crossing over a mine. As mentioned before the use of modular robots is perfect for the design of task-specific demining

robots because of the similarities between the tasks. Since the complexity of a system raises the susceptibility to trouble exponential it is always better to keep devices as simple as possible and therefore to use simpler robots. Using smaller robots extends the operational time before re-fuelling or re-charging is necessary or at least prevents the use of bulky and heavy batteries or tanks. On the other side the second possibility requires an increased effort in communication between the robots in the swarm. If every robot is able to perform the whole demining process by itself the communication is reduced more or less to get each other out of the way and ensure to cover all the area. However the task-specific robots have to exchange a lot more of data. The detection robots must work together since they are equipped with different detection technologies. When they have found a mine they must signal it to the removal robots and they have to inform after done work the transportation robots. 3.1 Detection Robots The robots for the detection of landmines are probably the most simple of the three types. The basic composition of modules common for all three types has to be upgraded only with the detection system. There are several detection technologies in use respectively under development, none of them able to detect a mine alone by itself. The solution is to use two or more of these sensors simultaneously. The first logical step would be to attach different sensors on one robot. Since there are weight limits, and limits in the amount of available energy, this is probably not the best solution. Some of these technologies need strong power sources and some of

Fig. 2: Humanitarian Demining Robot Swarms [4]

them are relatively heavy constructed. These facts will not help to keep the weight of the robot low, so using for each type of sensor a single robot seems to be the better solution. The detection robots should communicate with each other, change data and coordinate their work. If one robot with one distinct detection technology has found a possible target, the area should be verified by all other detection technologies before any further action is started. Therefore the different technologies must be compatible to allow coordination. At least the data from the sensors should be assessed by the same software. Combining results from different mine detection technologies is not easy and demands special strategies. These so-called sensor fusion technologies are not only of concern for mine detection. Another important point is the power supply of the detection swarm. One possibility is to equip the detection robots with an autonomous power source. But this could complicate the recharging of the system. There would be need for extra docking stations and at the worst for each detection technology a different docking station. Using the power source of the mobile robot platform would only alter the recharging process with regard to the operational time. In exchange compatibility problems could occur. Each detection technology may need power in a different way. And additional a modular system insists a quick exchange of the modules. Each sensor has to be mounted on every platform in a fast and easy way. It should be some sort of ‘plug and play’. This cooperation during the development and design process of the modular robot system and landmine detection sensors is of greater concern than only for an appropriate modular interface. Some of these sensor systems are extremely sensible and may drop in performance in presence of distinct materials. Using these materials for parts of the robot system which has to carry the sensor technology has to be avoided. And many of the sensor techniques work by using radiation in some range of the electromagnetic spectrum. It has to be guaranteed that systems of the robot do not jam the sensor technology or the other way round. 3.2. Removal Robots The removal of landmines is probably the heaviest work during the whole demining process. This is clearly a matter of the type of soil in which the mines are buried. But generally this task needs the highest forces and therefore the employed system has to be more stiff and heavy constructed.

The removal robots have also the most various part to fulfil. While the detection robots only transport the detection technology and the transportation robots have to accomplish an advanced pick and place task, the removal robots have to, in case of buried mines, dig out a highly sensitive device, which must be handled extremely carefully, but at the same time applying relatively high forces to penetrate the soil. In addition the excavation of a mine is every time a different procedure. The main parameters which differ for each buried mine are the type and shape of the mine, the position relative to the surface and the type of soil in which the mine is buried. Some mechanical mine clearance devices used today can be compared in some ways with the considered removal robots. One difference is that the mechanical devices are brute and heavy compared to the robots. This means that it doesn’t matter if they trigger off an explosion, because the explosion can not damage them. Beside the fact that explosions are undesired because of the environmental hazard, every time damaging a robot would be very costly and delays the progress of the whole work. To avoid this, the excavation of a detected landmine must be carried out considerate rather than ‘digging until hit upon something’. Since the excavation is a complex task a dexterous robot arm with a high number of degrees of freedom is likely to be used. For the mine removal various end-effectors may be necessary. The robot arm can be equipped with a variety of standard tools which are similar to tools used for manual excavation. All forms of shovels are doubtless of interest to remove foremost close grained material. Grippers may used to sweep stones or other bigger obstacles. These tools are commercially available and well proven. Up to the present the most removal work performed at hazardous materials was executed teleoperated. For that the aid of sensors is mainly limited to force and torque sensors which ensure not to apply too high forces to the sensible object. But the whole process is controlled by an operator using video cameras to lead the tools. Using a robot for autonomous removal of landmines presupposes the usage of sensors to compensate the teleoperator. Two broad classes of sensing technologies support earthmoving automation. One class allows determining the state of the robot itself, the other class concerns perception of the environment around the earthmover. Local state is achieved by measuring displacements at the robots various joints. If the actuators are hydraulic cylinders the use of position transducers

would be a good choice. An alternative is to use joint resolvers, like potentiometers, directly at rotary joints. Another form of state estimation is to locate the robot arm with respect to some fixed coordinate frame. Many sensing modalities have been used including, GPS, inertial sensors and reflecting beacons. Successful estimation schemes combine several of these techniques. These sensors which are used to perceive the environment during the excavation work have interesting similarities to the sensors used to detect the landmines. In fact some of the excavation sensors mentioned above are almost identical with them. Therefore the idea to use the detection sensor also to support the removal work is self-evident. Some modifications of the software, filters or the bandwiths may enough to switch from detection mode to removal mode. Beside the possibility to install some of the detection modules on a removal robot there is also the opportunity to let a detection robot and a removal robot work side by side during the removal process. The detection robot can perceive the changes in the environment and the progress of the excavation and transmit the obtained data to the removal robot. With this information the removal robot can plan and execute the next steps. The variety of possible tools used for removal actions makes a tool changing system onboard a removal robot necessary. The system should be actuated by the same principle as the tools are actuated. That implies that a whole set of excavation tools should be actuated by the same principle. Employing more sophisticated removal tools like the presented waterjet system returns some problems for the use of only one robot arm. In that case the end effector has to be supplied with more than the standard tools need. The water supply of this device is generally a problem in mobile robotics. A simple solution could be the use of different removal robots. One type of robot unvarying equipped with such sophisticated removal devices and a general type of removal robot equipped with standard tools and a tool changing system. However to a tool changing system always belongs an appropriate magazine for the tools. In outdoor applications it is important that the tool magazine protects sensible parts of the tools against dirt. Especially the interfaces for power and information supply should be kept clean to ensure a proper passing on when the tools are used. Therefore the magazine should be sealed off against the environment someway. 3.3 Transportation Robots The transportation seems to be quite simpler than the removal of a landmine. Basically the robot has to pick up the landmine, store it somewhere during the transportation and deliver it at the collection point.

An important decision in respect of the transportation robots is the number of mines the robots should be able to carry. Carrying only one mine would it make possible to use a rather simple robot. At the best it may possible to retrench the storing place for the landmine. The robot could pick up the mine with a gripper, lift it up somewhat above the ground and transport it to the collection area while holding it tight with the gripper. The use of a dexterous robot arm, like that one for the removal task, would be disproportionate. A simple 2 DOF lift onboard the mobile robot platform could be sufficient. On the other side the application of a transportation robot with the ability to carry more than one mine is in a manner useful too. Since transportation robots are likely to be rather slow this approach is much more timesaving. The volume of saved time depends on the amount and distribution of collection areas in proportion to the field of activity as well. But establishing lesser collection areas simplifies the further strategy for the disposal of the collected landmines. To give the robot the ability to transport more than one mine it must be equipped with some sort of storage device. On principle it would be of use to make the storage device of protective material to mitigate accidentally explosions. One possibility is to use a lockable storage device. But therefore the device must be designed with regard to a maximal allowed load of explosives. An explosion inside a locked container exceeding the maximal allowed load may be worse than without any protective measures. Fragments of the blasting container could damage the robot in addition. For this reason it would be better to use a container which is opened upwards. This guarantees a way out for the pressure wave in case of an accidental explosion. The placement of the mines inside the storage device is of special interest. Storing the mines one upon the other increases the danger of the explosion of a mine during the transportation process. Contrarily storing the mines side by side presupposes a large base of the storage device. Additional the design of the storage device and the placement of the mines inside should guarantee a simple load and unload process. A safe and space-saving storage device should not be installed at the cost of the lifting device. If for the loading and unloading a highly dexterous robot arm and a sophisticated end-effector are needed, it might be better to reduce the complexity of the storage device. An importing factor for the decision of using single or multi transport robots is the density of the minefield. If there are only few landmines per surface unit the application of single-mine transportation robots is more likely. In this case the work quota of

the detection robots is much higher compared to that of the removal and transportation robots. Therefore raising the working capacity of the transportation robots would not increase the overall efficiency perceptible. The features of the robot for these three tasks have to be quite different. For detection a light-weight robot only able to carry little load has to be developed. For removal the robot has to be more stiff and heavy constructed because removal requires force. The size of transportation robots depends on the number and kind of the mines to be transported. Another point of view which has to be taken into account is the time necessary for these operations. Detection is usually relatively fast and is not so time consuming than removal. According to some experiences the removing time is 3 to 5 times more than the detection time. Transportation time is also relatively small. Therefore it could be advantageous to use three different types of robots. Robots for detection, robots for removal and robots for transportation of the mines. One main disadvantage of this philosophy is if a robot of the swarm D (detection) has found or detected a mine it has to send a command to the host computer or to the other robots. The host computer or the other robots have to decide which of the robots of the swarm R (removal) is in the neighbourhood of this mine and not busy at that time with removal operations on another mine. If a robot of the swarm R is selected this robot gets usually wireless the position data and some other information about the place of the mine. The R robot is now moving to displace and start with the removal work. After the removal of the mine it has to place the mine on the ground in a distinct position. One of the transportation robots (T) have to pick up the mines and has to carry it to a collecting place. 4. Realisation Today we are in the position to develop robots of all three types mainly using commercially available mobile platforms. As pointed out earlier it is not economically feasible to develop so-called single purpose robots for each of these three types. A good approach could be a kind of a tool kit [5] of mobile robots consisting of a mobile platform and different equipments and tools compatible in hard- and software. A good approach could be to have two platforms, one with wheels or chains and one walking platform. According to the types of mines as well as the surface of the minefield these platforms could be equipped with necessary tools in a very short time. Usually the mobile robots of both types available today are moving relatively slow. Usual speeds for

wheeled and chained robots are between 0.5 and 0.7m/s, walking robots are usually much more slower. This could be a disadvantage concerning the demining time but from the viewpoint of control and path planning it is much more easier to work with such slow robots. We have in that case the usual problem of path planning of robots in a changing environment. Usually in a minefield we have fixed obstacles like trees, rocks, buildings as well as moving obstacles usually the robots of the own or other swarms. 5. Summary This approach is similar to “Multi Agent Systems – MAS” [3]. These systems are very well known in software engineering since more than 20 years. In the last years there are some works related to the application in production automation. A MAS consists of a number of intelligent, co-operative and communicative hardware agents e.g. robots getting a common task. Because of the intelligence they are able to divide the whole task in subtasks as long as at least one of the agent is able to fulfill one subtask. Repeating this procedure yields to the solution of the common task. Newest research goes in the direction of MMAS – Multiple Multi Agent Systems – different MAS are involved for the solution of a complex task. In a mid or long term perspective it might be possible to develop “ Humantarian Demining Multi Agent Systems – HDMAS ” consisting of a number of such robots or agents [3]. Robot swarms or HDMAS for demining are currently only a vision but will be reality in the nearest future. 6. Literature [1] Baudoin, Y. et.al. (2000): “Humanitarian

Demining: Sensory and Robotics”. Proceedings of the IMEKO World Congress 2000, Vienna, Vol. XI, p. 241 – 251.

[2] Kopacek, P. (1999): Preprints of the IARP Workshop on “Humanitarian Demining”, Harare, November 1999.

[3] Kopacek, P. (2000): “SWIIS – An Important Expression of IFAC’s Commitment to Social responsibility”, Preprints of the IFAC Workshop on Supplemental Ways for Improving International Stability – SWIIS 2000, May 2000, Ohrid, Macedonia.

[4] Kopacek, P. (2002): “Demining Robots – a tool for International Stability”. Proceedings of the 16th IFAC World Congress, Barcelona, July 2002

[5] Shivarov, N. (2001): “A tool kit for modular, intelligent, mobile robots”. PhD. Thesis, Vienna University of Technology, 2001.

ON-BOARD DEMINING MANIPULATOR

Aníbal T. de Almeida 1, Lino Marques 1, Michael Rachkov 2, Valery Gradetsky 3

1 Institute of Systems and Robotics, University of Coimbra, Polo II, 3030-290, Coimbra, Portugal 2 Moscow State Industrial University, ul. Avtozavodskaya 16, Moscow, 109280, Russia

3 Institute for Problems in Mechanics, Russian Academy of Sciences, Vernadskogo av., 101, Moscow, Russia

Abstract The paper describes the modelling of an on-board demining manipulator that contains a pneumatic drive, an infrared mine detector and a mine neutralizator. The infrared mine detector identifies the mine position in the scanning mode of the manipulator and gives a control signal to the input of the manipulator drive control unit for accurate positioning of the neutralizator above the detected mine. A problem of the optimal manipulator positioning in the sense of the control energy consumption minimization is solved. Modelling results of the infrared detector mine searching and of the neutralizator positioning by means of a pneumatic manipulator are presented. Keywords : modelling, demining manipulator, pneumatic drive, optimal positioning, infrared mine detector. 1. Introduction If the cost of the mine removing would be about the cost of the mine, the main advantages of using the mine would disappear. Automation of demining operations can carry out this task with substantially reduced costs due to the simple design of the demining system [1]. Automation of demining can be performed by means of autonomous robots [2] equipped with a mine detection block and a mine neutralizator. A robot manipulator carries out searching trajectories of the detection block and positioning of the mine neutralizator. For autonomous robots high payload-to-weight ratio of the manipulator is important. Pneumatic manipulators have such a possibility, compared to electric driven manipulators. Another desirable characteristic for autonomous systems is the minimization of the energy consumption of the on-board supply unit. This demands applying an optimal feedback control of the manipulator motion. It was concluded in [3] that the third-order control provides a practical choice for effective control of industrial pneumatic manipulators. The demining manipulator should fulfill the searching motion of the mine detector and the positioning of the demining equipment. Modelling of the infrared (IR) mine detector searching and the mine neutralizator positioning is presented.

2. System description The demining system consists of the manipulator that is installed on a mobile robot (Fig. 1). The end-effector of the manipulator contains the mine detector and the mine neutralizator.

1

2

34 5

6

8

9

10

7

Fig. 1. General diagram of the system

1 – mobile robot, 2 – wheels, 3 – manipulator, 4-manipulator drive, 5 – mine detector, 6 – mine

neutralizator, 7 – control and supply block, 8 – scanning trajectory, 9 – mine, 10 - neutralizator positioning

trajectory The mine detector performs scanning trajectories by means of the manipulator during robot motion across a minefield. After a mine is detected, the manipulator should perform the neutralizator positioning trajectory to place it over the detected mine. The neutralizator is based on laser heating of the mine until the explosive filler ignites and starts to burn. If the mine has a metal case, the heat is conducted through the case and target irradiation is continued until the temperature of the inside wall and the temperature of the explosive filler exceeds its combustion temperature [4]. If it is a plastic case, the case is irradiated until it has been penetrated and the explosive filler is ignited, either directly from the laser radiation or from the flames burning the plastic case. A functional diagram of the system is shown in Fig. 2. The mine detector provides information about the mine angle position in the scanning mode of the manipulator.

Controlunit

Manipulatordrive Manipulator

Mineneutralizator

Minedetector

Mine

Fig. 2. Functional diagram of the system

This information goes by a feedback loop to the control unit and changes the scanning mode (180° rotation) to the positioning mode (rotation to the detected area) of the neutralizator. The manipulator drive performs positioning of the neutralizator according to the given angle. The angle is measured relatively to the central axis of the manipulator taking into account the design parameters of the neutralizator and its connection to the manipulator. 3. Mine detector modelling The mine sensing is based on an infrared image analysis obtained during microwave soil heating and posterior cooling. The detector prototype contains a microwave klystron emitting 1 kW power at the frequency of 2.45 GHz and two infrared sensors sensitive in the range of 8-14 µm. Depending on the soil dielectric properties, the emitted radiation can be absorbed, reflected or transmitted through. Common plastic materials are transmissive, metals reflect the microwaves, and wet soil absorbs and converts the radiation to heat. Using this sensor, it is possible to image thermal gradients in the soil surface and detect different rates of temperature changes depending on the soil content [5]. The mine detector uses temperature gradients sensed over a homogeneous soil surface containing a plastic mine. According to the electromagnetic theory, a plane wave propagating in a lossy dielectric can be expressed by [6]:

0 0z z j zE E e E e eγ α β− − −= = (1)

where z is the propagation direction, E0 is the electric field in position z = 0, and α and β are attenuation and phase constants for the material in which the wave propagates. The propagation constant γ can be expressed by the following equation

( )j jγ ωµ σ ωε= + (2)

where 0rε ε ε= and 0rµ µ µ= are the dielectric

permittivity and the magnetic permeability of the material expressed relatively to the permittivity 0ε

and permeability 0µ in free space, σ is the material

conductivity and ω is the angular frequency of the wave. If σ is much bigger than ωε , the medium can be considered as a perfect conductor, if σ is much smaller than ωε the medium can be considered as a perfect dielectric. When a planar electromagnetic wave propagates into a soil surface, a part of it will be refracted into the soil and the other part will be reflected to the air (Fig. 3).

rEiE

tE

θi

θr

θt

µ1

µ2

Z

Fig. 3. Electromagnetic wave reflection and refraction According to the Snell’s law, reflection and refraction angles can be expressed by the following equations:

i rθ θ= (3)

sinsin

i

t

nθθ

= (4)

where iθ , rθ and tθ are incident, reflection and

refraction angles respectively, n – refraction coefficient. The ratio between the reflected (Er) and the incident electric field (Ei) is the reflection coefficient. Depending on the type of wave polarization used (horizontal or vertical), this ratio can be expressed by Rh or Rv:

2

2

cos sin

cos sini r ir

hi i r i

ER

E

θ ε θ

θ ε θ

− −= =

+ − (5)

2

2

cos sin

cos sinr i r ir

vi r i r i

ER

E

ε θ ε θ

ε θ ε θ

− −= =

+ − (6)

These expressions are useful to determine optimum incident angle in order to maximize the energy propagated into the second medium. The electric field can be related with the transmitted power using the Poynting vector and Maxwell laws [7]. In the far field, the electromagnetic power P can be expressed by:

212

P E Aη

= ⋅ ⋅ , [W] (7)

where η is the impedance seen by the wave (η = 120π in free space) and A is the soil area enclosed by the valve radiation (in our case A = 0.1225 m2). The heat generated in an elemental volume of material by a microwave electric field depends mainly on the frequency and on the dielectric properties of the material [8]. The power Pv absorbed per unit of volume can be calculated through the following equation,

2 tanPv Eω ε δ= ⋅ ⋅ ⋅ , [W/m3] (8) where ω is the angular frequency, E is the absolute value of the electrical field, ε is the material permittivity, and ( )tan δ is the tangent of losses in

the medium that can be expressed by the following equation [9]:

''

tan'

ωε σδ

ωε+

= (9)

where 'ε and ''ε are real and complex parts of the permittivity ( )' '' ' 1 tanj jε ε ε ε δ= + = − (10)

A computational implementation of the model was done by Matlab [10]. It calculates the microwave attenuation and the power absorbed from the valve output, above the ground, up to a specified depth in the ground. Vertical polarity is used by default in the implementation. The user can choose to visualize the power or the electrical field and can model the soil with or without a mine. It is also possible to visualize the power absorbed by the soil and by the landmine. A plastic landmine with rε = 2.3 and

( ) 4tan 0.66 10δ −= × [11] was considered. It is

assumed that the microwaves do not suffer attenuation while passing through the mines, since these are constituted, in its bigger part, by plastic material. The model includes a value for the dielectric constant of the ground. In general case, the ground is an anisotropic medium whose properties are changed with the frequency, moisture content and temperature [12]. The model contains the following parameters: f = 2.45 GHz, common frequency of

microwave heating systems, Pt = 1000 W, power emitted by the microwave klystron, depth = 0.5 m,

rε = 10, typical relative permittivity for sand,

conductivity = 10 mS/m, idem, iθ = 65º. The

reflection coefficient of 0.1649 and the optimal incidence angle of 72º were used in the model. The following graphics (Fig. 4 – 7) show the electrical field and absorbed power without the mine and with the mine of 5 cm height, as a function of depth.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.51600

1700

1800

1900

2000

2100

2200

2300

Depth [m]

Ele

ctri

c fie

ld [

V/m

]

Fig. 4. Electric field (without mine)

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.52.5

3

3.5

4

4.5

5

5.5x 10

4

Fig. 5. Absorbed power (without mine)

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.51700

1800

1900

2000

2100

2200

2300

Depth [m]

Ele

ctri

c fie

ld [

V/m

]

Fig. 6. Electric field (with mine)

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.50

1

2

3

4

5

6x 104

Fig. 7. Absorbed power (with mine)

An analysis of the graphics shows that the electric field does not suffer attenuation through the mine and, therefore, the mine absorbs almost no power. The general equation for mass and heat transfer, particularly in non-homogeneous materials is very complex. The general equation for heat transfer, including the term of convection and the volumetric source of heat Pv can be written as

0

Tc T q Pv

tρ υ

∂ + ⋅∇ = − ∇ ⋅ + ∂ (11)

where T is the temperature, q is the vector of total heat flux, ρ0 is the material density and υ is the speed vector for the heat transfer fluid. The term ( )pυ∇

relative to compressible fluids at a pressure p was ignored in the above equation. If the material is submitted to radiation, then the heat flux q should include the radiation term qR and the conduction term qc: c R e Rq q q k T q= + = − ∇ + (12)

where k e is the effective thermal conductivity. The volumetric heating can be a consequence of ohmic heating by an electric current flow in the material, surface currents generated by induction heating, or reorientation of electric dipoles due to dielectric heating. The modelling of a volume heating by microwave radiation can be done using the Partial Differential Equations (PDE) toolbox of Matlab. This toolbox provides a graphical interface that allows modelling the shape and properties of the materials to be studied. The following equation was calculated over the mesh of finite elements generated by the toolbox:

( ) ( )0 extT

c k T Pv h T Tt

ρ∂

− ∇ ⋅ ∇ = + ⋅ −∂

(13)

where ρ0 is the material density, c is the specific heat, T is the temperature, Text is the external temperature, k is the conduction coefficient, Pv represents the source of heat and h is the heat transfer coefficient by convection. This example considers the absorbed power equal to power released by the source of heat. The first step to simulate the soil heating is to draw the geometry of the problem boundaries using the toolbox graphical interface. In Fig. 8, the model representation of the soil, the mine and the surrounding air is shown.

air

mine

Fig. 8. Definition of the problem boundary

4. Manipulator drive control unit The on-board manipulator of a length L and an end-effector with the load (detection block and neutralizator) of mass m, is actuated by double-acting pneumatic power cylinders by means a gear with a lever d. The considered drive system with pressure variation in pneumatic power cylinders [13] is described by non-linear differential equations of the third order

( )

gVRT

V

lPFp

fmL

lFp

+−=

−=

ϕ

ϕϕ

&&

&&&

2

22

(14)

where ϕ is the angular position of the manipulator gripper, p is the current pressure difference in pneumatic cylinder volumes, F is the cross section of the cylinder piston, R is the gas constant, T is the absolute temperature of working gas, V is the full volume of the pneumatic cylinder, p is the pressure inside the cylinder chambers in an equilibrium position of the cylinder pistons, g is the molar gas consumption in the pneumatic cylinder chambers,

( )ϕ&f is the summand taking into account the friction force of the drive system.

The force of inertia for rather large values of mass m considerably exceeds the friction force in the drive system. In this case, it is possible to transform the system (14) as follows

uxax

xxxax

+−=

==

1313

12

3131

&&

& (15)

where

ϕ&=1x , ϕ=2x , px =3 ,

213

2

mL

dFa = ,

V

dPFa

231 = , (16)

gVRT

u =

Thus, the phase coordinates of the system are the angular position, the angular velocity of the manipulator end-effector, and the pressure in pneumatic power cylinders. A control parameter is the gas consumption. The problem of minimization of the positioning coordinates of the system (15) and, simultaneously, of the control energy consumption should be solved. It is possible to solve the optimal control task by means of the following quadratic criterion

(17)

where r2, r3, ρ are adjustable coefficients. The optimal control solution of this task is:

)( 3211 pPPPuo −−−= − ϕϕρ & (18)

where P1, P2, and P3 are coefficients of the feedback control loop [13]. The system using the optimal control (18) is asymptotically stable. The control signal is used in the drive for the optimal positioning of the manipulator. The simulation and an experimental test with the industrial pneumatic manipulator Tsiklon [14] with the following numerical parameters: P = 60 N/cm2, d = 8 cm, m = 200 N, L = 80 cm, F = 133 cm2, V = 2790 cm3, r2 = 10, r3 = 1, ρ = 28 were done. Simulation and experimental results for the 30º angle positioning (Fig. 9a) and 90º angle positioning (Fig. 9b) show that end-effector trajectories tend to the set angle exponentially quickly.

Min

e an

gle

posi

tion,

gra

d

Time, s

0 0.5 1.0

10

20

30

1

2

a) 30º angle positioning

Min

e an

gle

posi

tion,

gra

d

Time, s

0 1 2 3 4

30

60

90

1

2

b) 90º angle positioning

Fig. 9. Simulation results of the manipulator positioning 1 - modelling trajectory, 2 - experimental trajectory

An accuracy difference between modelling trajectories and experimental trajectories is inside 15%. This difference is caused by the friction influence in the pneumatic elements of the real manipulator. The accuracy discrepancy decreases by increasing of the rotation angle value, therefore, the modelling assumptions correspond in a satisfactory manner to the real positioning process parameters. The described modelling method can be applied to manipulators with other design parameters for the double-acting pneumatic drive. 5. Conclusions Modelling of the demining manipulator is an important task while developing demining robots, in order to estimate the design and the dynamic parameters of the detection block and of the manipulator drive unit. This methodology appears well suited to achieve the effective functioning of the automatic demining system. The IR detector functioning is modelled in the mine searching two-phase mode. The obtained temperature distribution inside a volume of soil containing a

( )dtuxrxrI ∫∞

++=0

2233

222 ρ

plastic mine after exposing the workspace with microwaves, permits to generate information about the location of the mine. This information is used as an input signal for modelling of the position control of the mine neutralizator by means of the double-acting pneumatic manipulator drive. The problem of the optimal manipulator positioning in the sense of the control energy consumption minimization is solved. Simulation results with numerical design parameters of the industrial pneumatic manipulator show that end-effector trajectories tend to the set angle exponentially quickly. The comparison between the modelling and the experimental results shows that the modelling assumptions are well adapted to the real process and that the proposed technique provides effectiveness in the system operation. 6. References [1] M. Rachkov, L. Marques, A. T. de Almeida

(2002). Automation of Demining , Textbook, University of Coimbra.

[2] A.T. de Almeida, O.K. Khatib (1998). Autonomous Robotics Systems , Springer.

[3] S.R. Pandian, Y. Hayakawa, Y. Kanazawa, Y. Kamoyama, S. Kawamura (1997). “Practical Design of a Sliding Mode Controller for Pneumatic Actuators”, ASME Journal of Dynamic Systems, Measurement and Control, 119, pp. 666 – 674.

[4] J. Yinon (1999). Forensic and environmental detection of explosives , Wiley.

[5] D. Noro, N. Sousa, L. Marques and A.T. de

Almeida (1999). “Active Detection of Antipersonnel Landmines by Infrared”, Annals of Electrotechnical Engineering Technology, Portuguese Engineering Society.

[6] D.M. Pozar (1998). Microwave Engineering , Second Edition, John Wiley & Sons.

[7] J. Griffiths (1987). Radio Wave Propagation and Antenas: an introduction, Prentice-Hall.

[8] N.E. Bengtsson and T. Ohlsson (1974). “Microwave Heating in the Food Industry”, Proc. of the IEEE, V. 62, 1.

[9] A.C. Metaxas (1996). Foundations of Electroheat – a unified approach, John Wiley & Sons.

[10] MATLAB (1992). High-Performance Numeric Computation and Visualization Software, The Math Works, Inc..

[11] J.E. Hipp (1974). “Soil Electromagnetic Parameters as Functions of Frequency, Soil Density, and Soil Moisture”, Proc. of the IEEE, V. 62, 1.

[12] M.T. Hallikainen, et al. (1985). “Microwave Dielectric Behavior of Wet Soil: Empirical Models and Experimental Observations”, IEEE Transactions on Geoscience and Remote Sensing , V. GE-23, 1.

[13] M. Rachkov (2000). “Simulation of an observer for a pneumatic manipulator control”, Proc. of the Int. Conf. on Identification and Control of Systems , Moscow, pp. 115-121.

[14] M. Rachkov (2001). “Optimal control of the pneumatic manipulator”, Control theory and systems , Nauka, 3, pp. 155-160.

SIMULATION AND CONTROL OF HIGH MOBILITY ROVERS FOR ROUGH

TERRAINS EXPLORATION

Ch. Grand, F. Ben Amar, F. Plumet and Ph. Bidaud

Laboratoire de Robotique de Paris, Universite de Paris VI

18, route du Panorama, 92265 Fontenay-Aux-Roses, France

E-mail: christophe.grand@robot.jussieu.fr

Abstract

This paper reviews works performed in our lab con-cerning the optimisation of locomotion performancesfor high mobility rovers evolving on rough terrains.Design and control of redundantly actuated locomo-tion systems are investigated in the aim to improvethe field of accessible terrain for autonomous explo-ration. Main targets of such studies concern the offroad locomotion in difficult environment like in vol-canic or planetary exploration.

keyword: hybrid locomotion, reconfiguration, forcebalance, stability, wheel-ground interaction.

1 Introduction

Design and control of high mobility rovers is a possi-ble way to enhance the field of accessible terrains forautonomous or teleoperated unmanned ground vehicle[1, 2]. Actually, this kind of vehicle is able to performcomplex motion like hybrid locomotion, and offers re-configuration and terrain adaptation capabilities.

Figure 1: the mini-rover prototype

This paper presents our research activities that con-cern: the rover-soil interaction model used for trac-tion control and its in-situ identification procedure,the simulator used for evaluation of the rover loco-motion behaviour on soft soil, and control algorithms

that enhance the rover locomotion performances. Foreach topic, the preliminary results are presented.

Hybrid locomotion systems take both advantages ofwheeled and legged vehicles that are first: the pay-load, the velocity and energy consumption, and sec-ondly: the stability, the high adaptability and obsta-cles clearance capabilities. We have developed suchkind of hybrid locomotion system that is the mini-rover prototype shown in figure 1.

It is a high mobility redundantly actuated vehicleand it is approximately 40 cm long and weights 10kg. It has four legs each combining a 2 DOF suspen-sion mechanism associated with a steering and drivenwheel. Thus, this 16 DOF system inherits advantagesof both legged and wheeled vehicle. The mini-rover isalso equipped with: two inclinometers to get informa-tion on platform orientation and a 3 components forcesensor on each leg to measure contact forces. The16 actuated mobilities provide the system with theability: to permanently maintain the four wheels onthe ground during displacements on uneven surfaces,to increase ground clearance, to increase the stabilityand the traction by controlling force balancing throughthe reconfiguration of the mini-rover. Moreover, thiskinematics structure allows the use of secondary lo-comotion modes like the peristalsis mode (crawlingmotion), and high obstacle-clearing mode based on acoordinated wheel-leg motion.

Peristalsis locomotion performed by the LAMA1

rover shown in figure 2, have also been analysed. Thissystem is a Marsokhod-type mobile robot [3] which hassix actuated wheels connected on three axles. Thisaxles are connected through two actuated rotationaljoints. All these mobilities provide the system withthe ability to perform peristalsis motion that consist tomove the inertial center by controlling the coordinatedmotion of wheels and axles. A previous experimentalstudy [4] have shown that peristaltical motion is moreadapted to climb slopes on granular soil than purelyrolling motion. In the best case on granular soil, this

1LAMA is property of ALCATEL-ESPACE

1

mobile robot is able to climb over a 25 degrees slopesby the use of rolling motion whereas it can climb over a30 degrees slopes performing the persitaltical motion.

Figure 2: the LAMA rover.

2 Wheel-soil Interaction Model

Optimisation of the traction performances plays a fun-damental role in locomotion of vehicles on natural out-door terrains: sand, clay, mud, or snow ... The trac-tion control deals with the optimisation of the wheeltorque which is directly related to the tractive force.This control algorithm interests the analysis of the lon-gitudinal wheel slippage ratio which is defined as therelative difference between the ideal rolling velocityand the real velocity of the wheel center. This slip-page is necessary to develop a traction force and par-ticularly on soft soils[5]. The traction force dependsalso on soil parameters, normal force, contact geom-etry, wheel stiffness, and crampon geometry... In thecase of our mini-rover, the characterization of the in-teraction can be done in situ by locking 3 wheelegsand proceeding to a shearing (and/or a compression)test of the local soil by acting on the fourth wheeleg.These tests need to measure of the normal and tan-gential contact force, and the slippage ratio that canbe computed from the joint velocities of the actuatedwheeleg. This experiments is similarly done by usinga Scara manipulator with driven wheel associated toa 6-axis force sensor[6].

Figure 4a represents experimentally measured tan-gential force coefficient commonly called the drawbarpull coefficient which is equal to the difference betweenthe tractive force (thrust) and the rolling resistancedivided by the vertical load. The rolling resistanceis mainly due to soil compaction and wheel sinkage.The curve corresponding to zero slippage (s = 0) rep-resents the rolling resistance since the tractive forceis theoretically null when there is ideally rolling. Thedrawbar pull coefficient is given on figure 4b as a func-

tion of the slippage for different vertical loads. Theobtained curves could be then represented by analyt-ical relations (similarly to Bekker’s relations for rigidwheels) characterising the global wheel-ground inter-action. The gaps between the 3 curves represent therolling resistance which increases more quickly thanthe tractive force when the normal load increases be-cause the sinkage increases more quickly than the con-tact area. The knowledge of the wheel-ground interac-tion allows to define the most suitable torque for thetractive wheel which is controlled through the use ofcourant sensor on each motor.

Figure 3: Testbed of wheel-sand interaction with aScara manipulator.

This method has the advantages to be simple toimplement and to be based on few sensors and somepreliminary in situ tests of wheel-soil interaction. Thismethod assumes that the mechanical properties of thewheel-ground properties are constant per area andmust be completed to integrate rules for detecting soilproperties changes.

3 Simulation of the Locomotion on Uneven

Terrain

Preliminary design and locomotion algorithms areevaluated through simulation of the dynamics be-haviour of the vehicle evolving on uneven terrain. Toachieve these simulations, we have developed a sim-ulation tool that integrates the whole dynamics ofmultibody systems, the behaviour of soft soil andthe interactions between the soil and the locomotionorgans[7, 8]. The simulator is based on a model de-scribing the deformation of the ground submitted toexternal forces. Figure 5 shows an illustration of thissimulator.

Due to its highly generic modelling approach basedon the Lagrangian Multipliers method[9], this simula-tor allows to evaluate the dynamic behaviour of vari-ous mobile robot evolving on rigid surface or soft soil

2

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0 0.5 1 1.5 2 2.5 3 3.5

Dra

wba

r pu

ll co

effi

cien

t

time (s)

s=0.00s=0.05s=0.10

s=0.30s=0.70

(a)

-0.1

0

0.1

0.2

0.3

0.4

0.5

0 0.2 0.4 0.6 0.8 1

Dra

wba

r pu

ll co

effi

cien

t

slip ratio

Fz=10N Fz=20N Fz=30N

(b)Figure 4: (a) Tangential force coefficient of wheel-groundcontact for a constant normal force Fz = 30N and fordifferent slippage ratio, (b) drawbar pull as function ofthe slippage ratio for different values of vertical loads

like sand. The geophysical properties of the ground areexperimentally defined by a triaxial test performed ona sample of soil. A semi-empirical model is used tointroduce the reaction force between the multi-bodysystem and the ground.

The definition of the ground geometry is based onfrequency synthesis that allows to generate realistic ar-tificial terrain [10]. By observing natural forms, it wasestablished that landscape forms have an A/f p fre-quency spectra where A defines the roughness and prelates to the fractal dimension. So, by using the spa-

Figure 5: illustration of the simulator

tial inverse Fourier transform of this spectral signal,an altitudes map of considered terrain is computed.

4 Locomotion modes

The mini-rover previously presented in the introduc-tion is able to perform different locomotion modes.This paper is focusing on the rolling motion mode witha reconfiguration of the platform attitude (i.e. pitchand roll angle). The reconfiguration capabilities areexploited in the aim to investigate control algorithmsthat enhance both the rover stability and the globaltraction performance [11].

Figure 6: results on the stability control algorithms

The platform attitude control consider the orienta-tion of the platform frame given by the three conven-tional roll-pitch-yaw angles (φ,ψ,θ) and it is based onthe definition of a stability and force-balance criteria.Furthermore when the system is moving, the tangen-tial plane of wheel-ground contact is difficult to deter-mine from the force sensor measurements. So, we willassume that the contact planes stay horizontal, i.e.,the ground is represented instantaneously by four dis-crete horizontal planes with different altitudes. Lastas the mini-rover velocity is low enough, only quasi-static stability is considered.

The aim of the control algorithm is to reach the moststable configuration from the current rover state. Inthese conditions and by considering quasi-static anal-ysis of forces distribution, we can assume that therover stability is maximum when vertical componentof contact-forces are equal on each leg. It is well knownthat vertical contact-forces balance can be reachedby minimizing the projected distance, on horizontalplane, between the rover center of gravity (c.o.g) andthe geometric center of wheel-ground contacts. More-over, this criterion also optimizes the traction forcedistribution. Consequently, if the ground is locallyhomogeneous in terms of its physical properties, theglobal traction of the propulsion system is enhanced.

By taking these assumptions into account, con-straint equations that are compatible with the stabil-ity criterion are defined. Due to the particular design

3

of the mini-rover, the sideways force balance, in thefront view, is obtained by constraining the platformroll angle to zero. The second constraint concernsthe forces balance in the sagittal plane, on right andleft sides. It is reached when each side wheelbase areconstrained in such ways that the projection of thec.o.g is in the middle of wheel-ground contacts. Then,two other free parameters need to be constrained: theground clearance and the pitch angle. The groundclearance defined as the mean altitude of each wheelis constrained to reach a desired value. This value iscomputed from a global planning algorithm under ge-ometrical constraints. The pitch angle is set to zerothat is well adapted when sensors like laser telemeteror vision system are used.

1 3 5 7 9 11 13 15 17 190.3

0.4

0.5

0.6

0.7

0.8

active control (average value = 0.570)no control (average value = 0.493)

Stability margin

t(s)

Margin (rad)

Figure 7: results of the stability control algorithms

Then, the control algorithm consist to compute themini-rover configuration (i.e. the position of each 2 doflegs) that is compatible with the previously definedconstraints. To perform the reconfiguration when thesystem is moving a velocity model based control havebeen developed. This consists to constraint the plat-form rotational velocity to reach the desired roll andpitch angle through a feedback control from inclinome-ter measures. And the vertical velocity of the platformis controlled to reach the desired ground clearance.

The control algorithm of the rover stability is de-scribed more accurately in [11]. This algorithm hasbeen studied by using the simulator presented in sec-tion 3, for a rover evolving on a rough terrain with avelocity of 0.3 m/s. The results in figure 7 shown thestability margin of the rover during is motion whenreconfiguration control is active or not. The stabilitymargin is computed from geometrical criterion definedin [12] and is given in degrees. These curves show thatthe mean stability of the system performing stabilitycontrol is 17% greater than with a fixed configuration.The minimum stability value is 27o in the case wherestability control is used, and is 19o in the other case.

a)

b)

c)

M1M3

M2

M1M3

M2

M1M3

M2

(a)

−1 −0.5 0 0.5 1−1.8

−1.6

−1.4

−1.2

−1

−0.8

−0.6

−0.4

−0.2

0

0.2

(b)Figure 8: (a) peristaltical motion with LAMA rover[4]

(b)contact forces distribution during peristalsislocomotion.

This represents an enhancement of the minimum sta-bility margin about 40%.

Peristalsis locomotion is also analysed through sim-ulations and experiments. These analysis are per-formed on the LAMA rover: we use its dynam-ics model for simulation of peristalsis locomotion onsandy soil[8]. Peristaltical motion improve the globaltraction efficiency on granular soil like sand. The sim-ulation shows that the tangential contact forces thatprovide the traction forces are more homogenously dis-tributed in the case of the peristaltical motion thanin the purely rolling motion. And all the simulationsshow that the maximum slope such rover can climbover is increased by 17% when using the peristalticalmotion. Experiments are also performing on the mini-rover in the aim to analyse for such kind of rover kine-matic this locomotion mode and how it can improvethe locomotion performance on soil made in granularmedia.

4

5 Conclusion

A global view of the research activities developed atthe LRP that concern the locomotion of high mobilityrover have been presented. The main objective of suchresearch is to provide systems that are able to adapttheir behaviour by them self, according to the locomo-tion environment variations in terms of its geometricaland physical properties. A model of the wheel-groundinteractions that allows to perform a traction controlof the rover from few sensor measures and a in-situcharacterisation test have been proposed. A simula-tion tools used to analyse more precisely the dynamicsbehaviour of such complex rover evolving on complexsoil has been described. And last, algorithms for thecontrol of the hybrid locomotion modes and their per-formances have been analysed. All these analysis arethe necessary preliminary studies for a more challeng-ing research on the global optimisation of the locomo-tion on uneven and complex natural terrain.

References

[1] K.Iagnemma, A.Rzepniewskia, S.Dubowsky,P.Pirjanianb, T.Huntsbergerb, and P.Schenker,“Mobile robot kinematic reconfigurability forrough-terrain,” in Proceedings SPIE’s Interna-

tional Symposium on Intelligent Systems and

Advanced Manufacturing, August 2000.

[2] A.Halme, I.Leppanen, S.Salmi, and S.Ylonen,“Hybrid locomotion of a wheel-legged machine,”in Proceeding CLAWAR’s of the International

Conference on Climbing and Walking Machine,(Madrid), 2000.

[3] A. Kemurdjian, “Planet rover as an object of theengineering design work,” in Proceedings ICRA’s

of the International Conference on Robotics and

Automation, (Belgium), pp. 140–145, 1998.

[4] G.Andrade, F.BenAmar, Ph.Bidaud, andR.Chatila, “Modeling wheel-sand interactionfor optimization of a rolling-peristaltic motionof a marsokhod robot,” in Proceedings IROS’s

International Conference on Intelligent Robots

and Systems, pp. 576–581, 1998.

[5] M. Bekker, Introduction to terrain-vehicle sys-

tems. The University of Michigan Press, 1969.

[6] J. Wong, Terramechanics and off-road Vehicles.Elsevier, 1989.

[7] C. Grand, F. B. Amar, P. Bidaud, and G. An-drade, “A simulation system for behaviour eval-uation of off-rooad mobile robots,” in Proceed-

ings CLAWAR’s of the International Conference

on Climbing and Walking Robots, (Germany),pp. 307–314, 2001.

[8] G. A. Barroso, Modelisation et Adaptation du

Mouvement de Robot Tout-Terrain. PhD thesis,Universite de Paris VI, 2000.

[9] P. Nikravesh, Computer-Aided Analysis of Me-

chanical Systems. Prentice-Hall, 1988.

[10] B.Mandelbrot, The fractal geometry of nature.W.H. Freeman, 1982.

[11] Ch.Grand, F. Amar, F.Plumet, and Ph.Bidaud,“Stability control of a wheel-legged mini-rover,”in Proceedings CLAWAR’s, 5th International

Conference on Climbing and Walking Robots,pp. 323–329, 2002.

[12] E. Papadopoulos and D. Rey, “A new mesureof tipover stability for mobile manipulators,” inProceedings ICRA’s of the International Confer-

ence on Robotics and Automation, pp. 3111–3116,1996.

5

COMPUTER VISION, AN HELP FOR DEMINING

Y. Caron, P. Makris and N. Vincent

Laboratoire d'InformatiqueUniversité de Tours64, av. Jean Portalis

37200 Tours – Francevincent@univ-tours.fr

Abstract

The problem of the deminization of large fields isalways thought of by means of physics solutions thatare relying on physical properties of the mines. Herewe come to the problem by analyzing the landscapeimages. Not all types of mines can be detected insuch a way but some can. We are developing anoriginal approach that makes some distinctionbetween natural and man-made objects using Zipflaw. This unables to point out a man-made elementwithin a natural scenery. The method is quiteindependent from the type of material the mine ismade of. The method could be coupled with someothers as a complement.

keywords : Segmentation, man-made objects, Zipflaw, object discrimination, natural scenery, texturecharacterization.

1. Introduction

Demining is a problem with the greatest importance,specially if considered from a humanitarian point ofview. Most of the land mines are hidden under theground earth but some of them may also be placed onthe surface of the ground as an natural object that iswell integrated in the environment. These mines canbe visually seen but as no one has any reason to payattention to them they can hurt as well if it happensthey are touched in any way. We are most interestedin these kinds of mines.

Most often, mine detection is achieved by use ofmechanical ways that rely on the properties of thematerials that are used when manufacturing themines. The emphasis is put on the sensors that can bedeveloped to achieve the detection. The detection ispossible when the mine is placed under the groundsurface as well as when it is visible. The variety ofused material increases the difficulty of the problem.Generally the sensors must be placed no too far fromthe mine and this leads to methods that need a nearpresence of human deminers and so the use of somerobots is necessary to ensure the security of thedemining process.

In this paper, we limit ourselves to the problem ofmines that are actually visible, that is to say the lightray comes from the object to the eye, but they are insuch a position that we can say they are hidden. Thisis a similar problem to the eggs that are hidden in thegarden for young children on the first of April. Herewe are proposing an approach that relies on computervision. Of course it is an absolute solution but itcould be seen as a complementary help to take adecision. It could also give some indices in order todetect some mines that are made of various newmaterials always more and more difficult to bedetected.

Of course computer vision is used in many otherways in the demining process as the sensors that areused can lead to data that are represented by imageson which the behavior of the sensor appears in adifferent way as the other elements of the observationfield or in a particular way that is known from thematerial that is looked for.

First we will explain the originality of our approachand how the mine can be considered as an extra pointin the landscape. Then the principle of Zipf law willbe recalled as it is the tool our method is relying on.Finally we present the method we are proposing andsome results.

2. Characterization of mines

Of course, mines are all man-made objects and theyare placed in some environment that most often isnatural surroundings. The aim of the mine is to lookas natural as possible, to take a natural place so that itdoes not seem quaint to the pedestrian in order toincrease the probability to hurt him. They are chosenin order to be adapted to the surrounding as achameleon does, changing shape or color accordingto the neighborhood.

Nevertheless they are built by machines with verysmooth shapes and their colors are very uniform. Infact no way is affordable to achieve object with ascomplex a shape or a color as is encountered in

nature. Then, we are proposing to seek for moreuniform zones in an image. The uniformity can beoccurring in the shape as well as in the colors.

One classical approach would be to extract all thedifferent objects present in the image. The use ofclassical edge detection methods is not efficientbecause most often too many contours are extracted.Then the interpretation is not achievable in areasonable time as no a priori information isavailable on the structure of the landscape. If we onlyextract most important contour, then too manyomissions would occur and information would lackto interpret the scene. The region detection methodsare also difficult to apply for the same reason. Eitherthe properties that characterize the object are toorestrictive or too cool. Then we obtain either toomany objects or too few. In any case theinterpretation would lack fiability.

Then, we are proposing to apply Zipf law to imagesin order to extract particular zones in images. Thesezones must be rather small within the image and theymust differ from the structural point of view from thesurrounding. Let us now precise Zipf law.

3. Zipf law applied to images

Zipf law states the distribution of patterns in a signalis not random. The relation between the frequencyand the rank of the patterns when sorted in thedecreasing order follows a power law. The n-tuplesof symbols are not randomly distributed. A powerlaw appears when the frequencies Nσ(i) of differentpatterns are ordered in decreasing order. This isdescribed by the expression:

Nσ(i) = k.ia (1)

Where Nσ(i) is the frequency of each symbol, i is therank of the symbol and k and a are constant. The “a”parameter characterizes the power law.

According to Mandelbrot, power laws characterizenatural configurations. Since Zipf’s law is a powerlaw, it can be used to model the distribution ofpatterns in an image. The a constant woulddistinguish between natural and man-made sceneries.

The method has mostly been used for 1D signal.Here we propose to use it in 2D images. The 2Dtopology is much more complex than 1D topology.The shape of a mask has to be chosen in order toadapt to the topology. Here we use 3x3 mask. In graylevel images (256 gray levels) too many patterns canbe realized, then no pattern has reasonableprobability to occur. Thanks to a coding phase thenumber of possible patterns has to be decreased. Weuse the general rank method that allows coding thepixels inside the mask using a 9 symbol alphabet.Then uniform region, what ever their gray level is, iscoded in the same way. Only differences between

gray levels are coded within the mask. In figure 1 isshown how the code process works.

255 210 210 5 4 425 2 34 1 0 240 2 40 3 0 3Figure 1 : Image pattern in grayscale level (left) and

associated general rank coding (right)

In order to visualize the behavior of different types ofimages a graph can be associated to the image. Itshows the variation of the number of patterns codedwith the general rank method with respect to the rankof the pattern they are all sorted in the decreasingorder. In figure 2 is indicated two different imagesand the corresponding graphs.

(a)

(b)

(a) (b)Figure 2. Images and associated graphs of natural (a) and

artificial (b) environments

As it happens, the experimentation of Zipf law onvarious images shows the law is holding for images.Besides, The images and associated Zipf graphspresented in figure 2 tend to show the nature of animage can be pointed out from the shape of thegraph. Thanks to the parameter values of the adjusted

law, it is then possible to discriminate between man-made images and natural ones.

Nevertheless, the problem is not to classify the imagein a global way but the problem is to detectlandmines that are just a part of the image. In nextsection we are to present the detection process.

4. The detection process

We can imagine a detector that comprises a videocamera that scans the environment. In fact thedistance between the camera and the potential mineis not known. Then the size of the object we arelooking for is not known and can vary. Surely theobject will not occupy the largest part of the imagebut rather some small part. We can state the size ofthe mine would not be too small, otherwise it will bedetected when the camera will come closer to themine. The position of the mine in the image is another problem. May be the image only comprise apart of the mine on one of the image border. In thesame was as for the size, we know that during thescanning, the mine would come in the middle of theimage. Then the experimental tests can be performedon images with a man-made object entirelypositioned within the image.

In order to find the land mines placed in the scenery,a global study is not the right thing to tackle theproblem. Rather the image is divided into many smallsub-images, let us name them frames. This is done inorder to have more homogeneous images to bestudied. For example we use most often 64 or 144frames that overlay the initial image. On each of theframes, the study is performed, that is to say the bestfitting Zipf law is computed. The previous parameterwe have defined are computed.

Figure 3. Use of sub-images

In the case of 144 frames, 144 area values associatedwith the local Zipf curve are computed. We calculatethe area that is defined on the Zipf graph, it is limitedby the x-axe and the graph itself. According to theinterpretation we have presented, the smaller the areais, the greater the probability is that the framecontains a man-made object. The achievement of thedetection is only possible when the object size is nottoo small compared to the size of the frame. From thecomparison of the sub-images parameters thelocation of a mine can be deduced.

In Figure 4 is shown in (a) a 3D representation of theparameter value according to the frames in a casewhere 256 frames have been considered. We obtain asurface with 256 vertices, and associated with animage that contains a man-made object, whereas in(b) the considered image does not contain anyartificial object. On the vertical axis, we haveindicated the opposite of the value of the area inorder to make apparent a potential extreme point forthe object location.

area parameter

(a)

area parameter

(b)

Figure 4. Zipf area associated with different images, weconsider an image in (a) that contains a man-made object

and in (b) that contain no man-made object.

In order to take the decision that a mine could bepresent some criteria are introduced, global and semi-global related to one sub-frame and the surroundingones in order to improve the discrimination power ofthe method.

Mostly two parameters are used, one is the area wehave presented in figure 4, the other is linked to thepower exponent and represent the complexity of theframe.

The area, illustrated in figure 5, is computedaccording to the following formula :

∑−

=

++ −+−=1n

1i

i1i1ii

2

)rr)(ff(A (2)

Where f is the frequency and r the rank of eachpattern. The sign is chosen here to obtain a peak onthe 3D surface when a man-made object is present onone frame. The object position is more visible, sincethe surface under Zipf’s curve is minimal when aman-made object is present.

Figure 5. interpretation of A parameter from Zipf graph.

The second parameter we have considered here, thepower law exponent, can be seen as a measure of thecurve’s average slope. The slope of the regressionline using least-square regression is given byfollowing formula (3) that is illustrated in figure 6.

∑ ∑

∑ ∑ ∑

= =

= = =

−

−=

n

1i

2n

1ii

2i

n

1i

n

1i

n

1iiiii

rrn

rfrfn

P(3)

Figure 6. Interpretation of P parameter from Zipf graph.

The detection is performed using those criteria. Ingeneral, on the graphical plot of the surface areaunder Zipf’s curve we can see a uniformity peak atthe object position and a large uniform zone in theupper part of the picture, which appear more uniformdue to the image field depth. The peak at the objectposition is not necessarily the absolute maximum inthe image, but it has the highest value compared tothe adjacent sub-frames.

Therefore we have defined a new detection criterionto improve the detection performances. It is theabsolute value of the difference between the surfaceunder Zipf’s curve for each sub-frame and theaverage value of that surface for the surrounding sub-

frames. If Aij is the surface under Zipf’s curve for thesub-frame (i,j), the difference will be:

−= ∑ ∑

+

−=

+

−=

1i

1ik

1j

1jlklij A

9

1AD (4)

With this criterion, the peak corresponding to thisobject is enhanced and the other high-uniformityzones are lowered, so the detection performanceswould be considerably better. In general the results ofthe average slope criterion are not as good as theresults given by the surface under Zipf’s curve.

In figure 7 we show different graphs associated witheach criterion. We can see the improvement as far thediscriminant power is concerned when the lastcriterion is used.

(a)

(b)

(c)

Figure 7. Detection results. image, surface under Zipf’scurve , Zipf’s curve average slope, area differences with

neighboring zones

In fact this method detects singular features whichdiffer from the surrounding background in respect topattern uniformity, whichever the difference may be.Then some object can also be detected if they are lessuniform than the background. This may occur as itwould be quite an alea that natural and man-madeobjects have the same structure. We are using anabsolute value in the formula that make nodistinction between more or less complexity in thedifferent zones.

5. Influence of a resolution change

In order to study the influence of an image resolutionon detection performances, the method was tested onresampled images. Detection tests were performed onseveral reduced versions of the same images. Adetection quality parameter is computed for thedifference of surfaces criterion. It is defined as theratio between the maximal value for the sub-framescontaining an object and the maximal value for thesub-frames without object. If this ratio is greater than1 the object is detected.

The results of these tests are extremely variableaccording to the images. In some types of images,reducing the image resolution will decrease detectionperformances and in some image it will enhancethese performances. There are even images on whichsuccessive reductions will have opposite effects. It istherefore impossible to predict the effect of an imageresolution change.

The influence of the number of sub-frames is alsovariable. For a better detection, it will therefore benecessary to perform several tests using differentimage resolutions in order to find the best one. Thecause of this phenomenon seems to be that thedifferent texture patterns of the image do not react toresolution changes in the same way, so their relativeuniformity will differ according to the imageresolution. The influence of the nature of thebackground would have to be studied more precisely.And the resolution could be adapted to differenttypes of landscape or environments.

6. Results

From our data base that comprises more than 80images, experiments have been performed and up to72% of the mines or other isolated man-made objetshave been detected. Our method is sensitive toobjects that cover 2.5% of the image. That gives anidea of the limit of the method that cannot too smallobjects.

Some filtering of the images could be performed inorder to enhance the different irregular zones. As aresult the use of filtered images reduces the numberof false peak detection by reducing the influence ofnatural uniform surfaces, which are not defined byvery precise edges and then are not enhanced by thepreprocessing of the image.

Besides the way the image is coded gives goodresults when highlighting is not uniform but themethod fail when the image is saturated or withproblem of depth fields.

7. Conclusion

As a whole, the method that we have presented isperforming quite well. It relies on the visualproperties of man-made objects and on themeasurement of some relative complexity of theinner structure of the elements that are present on anyimage.

Of course it cannot be used as a single method toachieve mine detection but it could be quite acomplementary approach of many others. Somefusion process can be performed between the resultsfrom Zipf law application and results from othermore physical approaches.

Besides, The same techniques could be developed tointerpret some images that are issued from differentsensors that are actually used in the detectors able tofind anti-personnel land mines.

References

[1] P. Beaver, S.M. Quirk, J.P. Sattler (1995).Object Identification in Greyscale Imageryusing Fractal Dimension, in M Novak, FractalReviews in the Natural and Applied Science,Chapman & Hall, London, pp. 63-73.

[2] D. Bi, J.P. Asselin de Beauville and M.Mraghni (1996). Spatial gray levels distributionbased unsupervised texture segmentation, 3rd

International Conference on SignalProcessing (ICSP96), Pekin (China).

[3] M. J. Carlotto and M. C. Stein (1990). AMethod for Searching Artificial Objects onPlanetary Surfaces, Journal of the BritishInterplanetary Society, vol.43.

[4] P.E. Forssen (1997). Detection of Man-madeObjects in Satellite Images. Thesis, Depart. ofElectrical Engineering Computer Vision,Linköpings University.

[5] B. B. Mandelbrot (1983). The FractalGeometry of Nature, W.H.Freeman, NewYork.

[6] D.M. Mark and P.B. Aronson (1984). Scale-Dependent Fractal Dimensions of TopographicSurfaces : an Empirical Investigation withApplications in Geomorphology and ComputerMapping, Mathematical Geology, Vol.16, n°7.

[7] V. K. Shettigara et al. (1995). Semi-AutomaticDetection and Extraction of Man-Made Objectsin Multispectral Aerial and Satellite Images. inAutomatic Extraction of Man-Made Objectsfrom Aerial and Space Images. Monte Verità.Birkhäuser Verlag Basel. pp. 63-72.

[8] N. Vincent, P. Makris and J. Brodier (2000).Compressed Image Quality and Zipf’s Law,Internatinal Conference on SignalProcessing, Beijin (Chine), pp. 1077-1084.

[9] G.K. Zipf (1949). Human behavior and theprinciple of "Least Effort" . Addison-Wesley,New York.

A COLOR CONSTANCY APPROACH FOR ILLUMINATION INVARIANT COLOR TARGET TRACKING

Geert De Cubber Hichem Sahli Hong Ping Eric Colon

Vrije Universiteit Brussel Royal Military Academy Pleinlaan 2 Avenue de la Renaissance 30

B-1050 Brussels-Belgium B-1000 Brussels-Belgium {gdcubber, hsahli}@etro.vub.ac.be eric.colon@rma.ac.be

Abstract In this work, we present a color target tracking

algorithm aimed at robot localization in varying illumination conditions. In our approach a colored target is put on the top of the robot and a fixed camera is used to detect and track the target. The 3D robot position can be estimated knowing the camera parameters after an analysis of the camera image. The general setup of this approach is sketched on Figure 1:

Many robotic agents use color vision to retrieve quality information about the environment. In this work, we present a visual servoing technique, where vision is the primary sensing modality and sensing is based upon the analysis of the perceived visual information. We describe how colored targets can be identified and how their position and motion can be estimated quickly and reliably. The visual servoing procedure is essentially a four-stage process, with color target identification, motion parameter estimation, target tracking and target position estimation. These individual parts add up to a global vision system enabling precise positioning for a demining robot. Mob t

A fixed camera

ile robo

Color target

Camera optical axis

Geometry center of the target

PC and DSP

Keywords : visual servoing, color constancy, robot positioning Figure 1: Setup for 3D robot localization

The first stage in the target tracking algorithm is the color target identification process, where a colored object that moves independently of the observer has to be found. Common image segmentation and object recognition were developed in the past to deal with this task. Yet, all of these algorithms have big problems as soon as the protected lab environment is left and tests are carried out in an outdoor environment where harsh and ever-changing illumination conditions cause great difficulties for the image-processing task, as a change in illumination will also change the perceived colors – or more generally the perceived image – of the environment. Numerous attempts have been made to solve this so-called “color constancy” problem and promising results have been shown before, as summarized in [8]. E. H. Land and J. J. McCaan were the first to tackle the problem with their retinex theory [7]. Others relied on finite-dimensional linear models [9][10][11][12][13], while also neural nets have been proposed as a solving technique [14]. However, these techniques commonly require hours of calculation time to process one non-synthetic image, which makes them totally unfit for real-time and real-world vision tasks, as is the case in the field of robotics. Here, we propose a color constancy technique used

1. Introduction The research work presented in this article fits in a global research effort to develop intelligent humanitarian demining robots. One of the tasks set up for these robots is to build precise maps of the inspected terrain with an indication of all the suspicious points where mines could be located. To do this, the robot must be equipped with a very performant localization system. However, classical absolute positioning sensors like GPS do not deliver the demanded precision in some cases. In these cases, the robot positioning problem can be solved using a single external camera as we present here. The use of computer vision for solving detection, tracking and positioning problems is a major research area in robotics [16][17][18][19][20]. Two approaches have been considered: (1) a fixed camera configuration, that is, the camera is fixed at a certain point in a general frame (the world coordinate system) [18], and (2) the eye-in-hand configuration, where the camera is fixed on the end-effector or mounted on a mobile robot [19][20]. For both cases, monocular [16][19][21] or stereo [18][22][23][24] vision systems have been investigated.

for real-time target identification under varying illumination conditions. Once the target object is identified, it can be tracked. Target-tracking refers to a method that enables a visual system to locate the target in its field of view using consecutive images. In our system the target is mounted on a mobile robot. The target tracking problem is regarded here as a camera control problem. We use the parameters estimated during the target detection step as well as the camera calibration parameters to control the camera’s motion (pan and tilt) and try to keep the target center coincident to the image center. Moreover, the distance to the target is estimated using a proportional scaling, so that a precise positioning of this target object, the robot system as a whole or just the end-effector can be performed. The rest of this paper is organized as follows: first, we will explain the used color constancy approach to achieve an illumination invariant color classification. Next, we show the working principles of the actual target tracking process. In paragraph 4, we will discuss the target localization procedure, which is most important for robotic applications. Finally, we present some results and conclusions. 2. Illumination invariant color classification Our approach is directly based upon the physical characteristics of color reflection. The main problem for the correct interpretation of a camera image is that the measured intensities are function of a huge number of parameters and most of them cannot be retrieved in any possible way due to their strong interconnectivity. The color of an object in the image is therefore more an appearance than a real material property. Nevertheless, color can be used to identify objects as long as the parameters which influence the formation of the perceived color are taken into account. To do this, we make use of the dichromatic reflection model, which was first introduced by Shafer in [1]:

( , , ). ( ). ( ). ( ). ( , , ). ( ). ( ). ( ).c b c b s c sk n i v e f r d k n i v e f r dλ λ

ρ λ λ λ λ λ λ λ= +∫ ∫ λ(1)

With: ρc: the measured intensity of channel c e(λ): the normalized light spectrum fc(λ): the cth channel sensor response function r(λ): the surface reflectance function kb: attenuation factor for the body reflectance ks: surface reflectance attenuation factor n : the normal to the surface patch

i : the direction of the illumination

v : the viewing direction Among the different color spaces, our choice went out to the l1-l2-l3-space, a color space which was originally introduced by Gevers and Smeulders in [6]

as a space that uniquely determines the direction of the triangular color in the RGB space. It poses an attractive alternative to the HSI space due to its computational simplicity. The space can be formulated as follows:

1

2

3

R Gl

R G R B G B

R Bl

R G R B G B

G Bl

R G R B G B

−=

− + − + −

−=

− + − + −

−=

− + − + −

(2)

In [15], Gevers and Smeulders prove that according to the dichromatic reflection theory, this space is invariant to highlights, viewing direction, surface orientation and illumination direction. This means that we can work with a simplified form of equation 1:

1 2 3 ( , ) ( , ). ( ). ( , ).l l l c bH x t e t f r x dλ

λ λ λ− − = ∫ λ

r

(3)

Equation 3 can be discretized by sampling over a number of wavelength bands. We chose to use a finite dimensional linear model with a limited amount of parameters and using 10 basis functions:

( , ) .( , ) .

e e

b r

e t B qr x B qλλ

=

= (4)

The columns of the N x Ne Be matrix and those of the N x Nr Br matrix represent the basis functions for the light spectrum and the reflectance spectrum respectively. The Ne element qe vector and the Nr element qr describe respectively the illuminant and the body reflectance spectrum. The problem with this representation is that the basis and sensor sensitivity functions are not well known. To avoid this difficulty, we use an approach similar to the one described in [4], which introduced a lighting and reflectance matrix, parameterized using 4 x Ne variables in a manner that is independent of basis functions and sensitivity functions. This leads to a general equation:

.T T

eh q σ= (5) With:

hT = (h1 h2 h3) σ = (σ1 σ2 σ3) a Ne x 3 matrix holding all the

reflection characteristics for a specific image point

In a learning phase, the algorithm learns the reflection characteristics of the object to be tracked. Small patches of images are accumulated over time while the material in question is subjected to a varying illumination. All intensity measurements are combined in an f x 3.p color measurement matrix H, while p is the number of pixels in the scene patch and f the number of frames sampled. If we sample for long enough, then eventually f will grow larger than p and the light spectrum matrix Q and the reflection

characteristics matrix S can be recovered by applying singular value decomposition on H. Thus, all factors in equation 6 can be calculated:

.H Q S= (6)

At this moment, the light spectrum distribution if the illuminant l is known p(qe|l) can be calculated. This can be done because Q is independent of the material. We use an Expectation Maximization (EM) clustering method [3] to derive the reflection distributions. This algorithm applies multivariate Gaussian mixture modeling with an unknown number of mixture components, so the number of clusters isn’t fixed on beforehand, which makes the classification very flexible. The result of this calculation is an NLS x Ne light spectrum matrix L and an Ne x 3 reflectance spectrum matrix R, with NLS the number of illuminant spectra distinguished by the EM algorithm. Now that we have estimates of the reflectance spectrum of the target object and now that we’ve obtained illuminant spectra corresponding to different lighting conditions, we want to correctly classify newly presented pixels as belonging to the target object or not, while keeping track of newly arising lighting conditions. We present a Bayesian solution to solve these problems. New scene properties are brought into the model based upon the Maximum A Posteriori (MAP) estimate of these parameters given the color measurements. When applying this classification, we search for the conditions that maximize p(o=oTarget, l, qe, σ|h) for any values of the lighting condition l, the illuminant spectrum qe and the reflectance spectrum of the target object σ, given the color measurement triplet h. The equation we want to solve is:

[ , ]

[ , , ] ( , , , | )e

e el q

o l q p o l q hargmax σ=

(7)

Using Bayes’ rule, it can be shown that:

( , , , | ) ( | , ). ( | ). ( ). ( )e e ep o l q h p h q p q l p l p oσ σ∝ (8)

The pixel classification procedure calculates the probability for each pixel and labels the pixel as belonging to the target object or not based upon the result. Using this theorem, the pixel classification is no longer performed directly based upon the pixels color value, as is classically done, but based upon the derived reflection characteristics. This method makes the detection process very robust and recovers the target shape, which enables the estimation of the target image size.

( , , , | )ep o l q hσ

During the actual tracking phase, the illumination model is continually updated using Bayesian reasoning. In this model updating stage, estimates for new lighting conditions and their corresponding illuminant spectra are calculated. It is this procedure that ensures the adaptive nature of the pixel classification process within the general target-

tracking program. The philosophy of this procedure is that we take a small patch from the target object, try to recover the spectrum of the illuminant shining on this part of the target object and update our model if necessary. This algorithm doesn’t need to run completely at every iteration, since there won’t be a new illumination condition with every new frame and only noteworthy changes in illumination will result in the model being updated, so there are a lot of exit conditions built into the process. The calculation of the new illumination condition itself can happen very rapidly, since we already know the reflectance spectrum matrix. After acquiring a nominal color triplet measurement hN, we can write:

1( ) .e new Nq N h R−= (9)

With Nnew the index of the rarest illumination condition within the L matrix, which will thus be replaced by the new lighting condition. After the pixel classification process, the target object will never be completely recognized, there will always be outlier pixels. This is shown on figure 6 in the results paragraph. To solve this problem, we use morphology filtering. During the color detection we create a corresponding binary image. For this binary image we use morphology filtering to do image segmentation. A square mask of 5 by 5 pixels is used as the structuring element in our application. Figures 7 and 8 show processed results. The white color shows the detected region after image segmentation and all the pixels in this region are considered as the detected pixel. The region is connected and represents the target shape very well. 3. Target tracking In fact, the target-tracking problem can be regarded as a visual servoing problem. In our system the target is mounted on a mobile robot. A calibrated camera fixed at the origin of the world frame is controlled through its pan (α) and tilt (β) angles to bring the target image center onto the image plane center. The camera zoom is also controlled to maintain a high signal-to-noise level. Figure 2 shows the camera control parameters which were defined.

Camera

Platform

Control

System

u

t

1α

1β

y1 y y’ P x p x’

c o x1

h

f

Figure 2: Camera control parameters

The proposed camera control method consists of two parallel processes: one process controls the pan/tilt camera platform in order to track the target; the other process uses a predictor to track the target in the image plane. Due to the fact that the robot moves with an unknown model, the servomotor-camera-target system is a time variant system. The target

motion model has to be identified in real time. Motion parameter estimation uses the visual input to estimate the dynamic properties of the target object. Initially, we set up a computational model based upon the theoretical aspects of the different components of the visual servoing system. This step comes basically down to perform a parameter identification for this theoretical model. In order to meet the system dynamic characteristic requirements we developed a two-phase control strategy. The first one is an initialization phase, in which the motion dynamics are estimated and during which the target is tracked with a PI regulator. The more interesting second phase control consists of a feedback control strategy shown in Figure 3. The plant is modeled as a dynamic system as shown in Figure 4.

1)( −Α−Ι⋅zΒ C

Β⋅⋅−Β CG

G z

Α C

1)( −Α−Ι⋅zΚ−

)(zΥ

+

+

+

-

Plant

Observer

Controller

)(zΧ)(zU

Figure 5: The Observer Based Full State Feedback Control

System

With: A: the plant system matrix given by [ ]10 ,aaB: the plant input matrix [0,1] C: the plant output matrix given by [ ]10,bb

G: the Kalman filter gain matrix

Plant

Observer

Control Strategy

yu

q

Identifier

K: the control gain matrix defined as the difference between the required and identified system parameters of characteristic functions

Using this Kalman-filter-based camera control strategy, it is possible to achieve smooth and stable camera movement, even when the target object undergoes shaky movements. The ability to estimate the target motion and to perform very rapid processing makes window-tracking possible. The proposed window tracking method reduces the image processing time and increases the signal-to-noise ratio significantly.

Figure 3: Feedback control strategy

4. Target size & distance estimation The visual servoing system presented here involves a method for estimating the target position, i.e. the quantitative description of where the target is with respect to the observer’s view. For our application, the similarity of the target shape and its projected image is used to estimate the camera/target distance.

Figure 4: Model of the plant

In our implementation a second order difference model is considered. The system functions are

The origin of world frame is set at the center of the camera. The camera platform is kept horizontal. Then, the position of the target can be described by 3 parameters: the horizontal angle, the vertical angle and the distance between camera and target. Angles are calculated using the pose of the camera and the orientation angles of the target image in the camera coordinate system. The distance between camera and target is estimated by simple similar triangle relationship of the real target size, the detected target image size and the effective camera focal length. The size of the target is estimated using circle and ellipse fitting procedures to more accurately measure the radius of the target object in the image plane. For the ellipse fitting, we used a very fast algorithm described in [5], whereas the circle fitting procedure is a much slower, but slightly more precise homemade algorithm.

( )( )

( )( ) ( )kukxkx

kakakxkx

+

−−

=

++

10

)()(10

11

2

1

102

1

( ) [ ] ( )( )

=

kxkx

kbkbky2

110 )()( (10)

Where: - ( is the state vector corresponding to the camera angles and angular velocity.

)

)

21, xx

- ( are the system parameters to be estimated.

1010 ,,, bbaa

These parameters are estimated using LSM method from a set of input image frames and camera control parameters. The feedback control strategy is implemented with system state vector estimation using Kalman filtering. The detail of the observer-based full-state-feedback control system configuration is shown in Figure 5.

5. Results Figure 9 shows the distance measurement errors

during indoor tests. As you can see, the errors do not exceed 0,12 meters.

Showing the results of the presented color target tracking approach and its illumination invariant features is kind of hard if the use of color is not allowed. Figure 6 shows a result of the pixel classification procedure. Remember this is still before the morphology filtering.

The Measurement Errors of Vision System

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

Mag

nitu

de (m

eter

s)

ErrorX 0.049 -0.007 0.057 0.091 0.11 0.087 0.053 0.064 0.0169 0.032

ErrorY -0.111 -0.108 -0.117 -0.052 -0.047 -0.048 -0.057 -0.053 -0.044 -0.093

1 2 3 4 5 6 7 8 9 10 11

Figure 9: Distance measurement error

Concerning the real-time capabilities, the classification algorithm takes about 60 ms to complete on a PC equipped with an Intel PIV 1.7GHz processor. When adding the 30 ms needed for morphology filtering and 10 ms for other tasks, we see the target-tracking program running at about 10 fps. This is adequate for every-day target tracking tasks, but not for high performance applications, so we still might want to improve the implementation a bit. The most processor-intensive process here is the management of the illumination maps.

Figure 6: Pixel classification

Figure 7 shows one scene of an outdoors static trial. The target detection function works well (detected pixels are painted white).

6. Conclusions We have shown a powerful set of algorithms, which were combined to form a universally useable system for automated target detection, tracking and position estimation, using a single and fairly simple pan/tilt camera. The Bayesian-based color constancy approach which was used ensures that this system can keep working, even in harsh illumination conditions. This research was specifically aimed at applicability in the field of robotics and due to its general structure it can also be used for a very wide range of applications. To conclude, we show a picture of the demining robot used for testing.

Figure 7: Outdoor trial Figure 8 shows an indoor scene during a trial on varying illumination. You can see that the image is very dark, because the lights were turned off, yet the target object is still detected almost entirely.

Figure 10: Demining robot

Figure 8: Extreme dark illumination conditions

7. References [1] S.A. Shafer, “Using color to separate reflection

components” (1985). COLOR research and application, Vol. 10, No. 4, pp. 210-218.

[2] P. Hong, H. Sahli, E. Colon, Y. Baudoin (2001).

“Visual Servoing for Robot Navigation”, Third International Conference on Climbing and Walking Robots: Clawar 2001, pp. 255 – 264, Karlsruhe, Germany.

[3] K. Zhang, L. Van Kempen, H. Sahli (2000).

“Finding maximum likelihood mixture density parameters via Expectation Maximization and its application to mine detection”, Vrije Universiteit Brussel.

[4] Y. Tsin, R. Collins, V. Ramesh, and T. Kanade

(2001). "Bayesian Color Constancy for Outdoor Object Recognition", IEEE Conference on Computer Vision and Pattern Recognition (CVPR'01).

[5] Markus Vincze, “Robust tracking of ellipses at

frame rate” (2001). The Journal of the Pattern Recognition Society, vol. 34, No. 2, pp. 487-498.

[6] T. Gevers, A.W.M. Smeulders (1999). “Color-

based object recognition,” Pattern Recognition, Vol. 32, pp. 453-464.

[7] E. H. Land and J. J. McCaan (1971). “Lightness

and the retinex theory”, Journal of the Optical Society of America, 61:1–11.

[8] K. Barnard, “Practical Colour Constancy” (1999).

PhD Thesis, Simon Frasier University, School of Computing.

[9] M. H. Brill and G. West, “Chromatic adaptation

and color constancy: A possible dichotomy” (1986). COLOR Research and Applications, 11:196–204.

[10] M. D’Zmura and P. Lennie, “Mechanisms of

color constancy” (1986). Journal of the Optical Society of America A, 3(10):1662–1672.

[11] B. V. Funt, M. S. Drew, and J. Ho, “Color

constancy from mutual reflection” (1991). International Journal of Computer Vision, 6(1):5–24.

[12] J. Ho, B. V. Funt, and M. S. Drew, “Separating a

color signal into illumination and surface spectral components” (1990). IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(10):966–977.

[13] L. T. Maloney and B. Wandell, “Color constancy: A method for recovering surface spectral reflectance” (1986). Journal of the Optical Society of America A, 3(1):29–33.

[14] B. Funt, V. Cardei, and K. Barnard, “Learning

Color Constancy” (1996). Proc. IS&T/SID Fourth Color Imaging Conference: Color Science, Systems and Applications, pp. 58-60.

[15] T. Gevers, H. Stokman “Reflectance Based Edge

Classification” (1999). Vision Interface, Trois-Rivières, Canada.

[16] A. Maruyama, M. Fujita, “Visual Feedback

Control of Rigid Body Motion Base on Dissipation Theoretical Approach” (1999). Proceedings of the 38th Conference on Decision & Control, pp. 4161-4166.

[17] E. Zergeroglu et al., “Robust Visual-Servo

Control of Robot Manipulators in the Presence of Uncertainty” (1999). Proceedings of the 38th Conference on Decision & Control, pp. 4137-4142.

[18] Ming Lei and Bijoy K. Ghosh, “Visually-Guided

Robotic Motion Tracking” (1992). Proceedings of Thirtieth Annual Allerton Conference on Communication, Control and Computing, pp. 712-721.

[19] Billibon H. Yoshimi and Peter K. Allen, “Active,

Uncalibrated Visual Servoing” (1994). IEEE International Conference on Robotics and Automation, pp. 156-161.

[20] Hong Zhang and James P. Ostrowski, “Visual

Servoing with Dynamics: Control of an Unmanned Blimp” (1999). IEEE International Conference on Robotics and Automation, pp. 618-623.

[21] Rafael Kelly, “Robust Asymptotically Stable

Visual Servoing of Planar Robots” (1996). IEEE Transactions on Robotics and Automation, Vol. 12, No. 5, pp. 759-766.

[22] Nicholas Ayache and Francis Lustman,

“Trinocular Stereo Vision for Robotics” (1991). IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 13, No. 1, pp. 73-85.

[23] Philip F. McLauchlan et al., “The Pipe-group

Architecture for Real-Time Active Vision” (1997). Real-Time Imaging, Vol. 3, pp. 319-330.

[24] James L. Crowley, “Dynamic Calibration of an

Active Stereo Head” (1993). ICCV93, Berlin.

TERRAIN ADAPTIVE SCANNING OF CONVENTIONAL MINEDETECTORS

Robert ChesneyYogadhish Das

Canadian Centre for Mine Action Technologies

Abstract

The Canadian Centre for Mine Action Technolo-gies1 (CCMAT) was established in 1998 with thegoal of developing technology for humanitariandemining. One component of the research pro-gram is the development of an articulated roboticscanner to allow conventional mine detectors tobe mechanically scanned in a manner similar tothe scan patterns of human operators.

The articulated robotic scanner (ARS) is a pur-pose built robotic arm designed to hold a minedetector payload (of up to 5 kg) and scan it overthe ground. The arm control system includes ter-rain measurement sensors to allow the detectorhead placement to be adapted to follow the ter-rain profile. The ARS design employs five degreesof freedom to allow a local work area of approx-imately 2m by 0.5m to be scanned. Within de-sign limits, the detector head can be “rolled” or“pitched” with respect to the horizontal plane,maintaining the head parallel to the local groundsurface, if desired. Coupled with the motion of ahost unmanned ground vehicle the ARS can scanany desired area; generating a spatially registeredmap of detector response.

The ARS design includes a second arm, with anadditional degree of freedom, to mount the terrainheight sensors. Height measurements are made bya combination of a scanning laser range finder andultrasonic distance measurement devices.

The paper briefly describes the development ofthe arm system, discusses the performance of thearm system in trials to date and touches on someof the system integration and data interpretationissues remaining to be solved before the approachcould be utilized in field operations.

Keywords: Demining, metal detector, areascan, robotic, tele-operation, humanitarian, vehi-cle mounted.

1. Background

A capability to automatically detect mines overlarge areas would make a significant contribution

1Box 4000 Medicine Hat, Alberta, CANADA

to mine action. Such a capability could be usedin area reduction2, actual mine clearance oper-ations or in quality assurance3. Numerous ve-hicle mounted systems have been developed tomeet this need[1, 2, 3], however, the focus of thiswork has largely been on detection of anti-vehicleland mines for military applications. Humanitar-ian demining operations face a significantly dif-ferent set of issues than military countermine op-erations. Primarily, in demining operations, thefocus is on the removal of all mines, including anti-personel mines, over a large area. In contrast,military operations often focus on providing safelanes for vehicle traffic, ignoring anti-personnelthreats and any area that isn’t immediately use-ful.

The ARS concept was originally developed un-der the auspices of the countermine research pro-gram of Defence R&D Canada[4] and was sub-sequently adopted by CCMAT based on the as-sessment of a scoping study [5] which included anexamination of the potential of scanning detec-tor systems in a purely humanitarian deminingcontext. The scanning system was aimed at du-plicating the complex scanning capabilities of ahuman operator, yet provide for computer basedprocessing of the detection information throughreal time measurement and control of the detec-tor head position. This led to the development ofthe current articulated robotic scanner or ARS.

The ARS development was conducted under aseries of contracts funded by the Canadian Cen-tre for Mine Action Technologies and the Cana-dian Department of National Defence. The armwas developed by Engineering Services Incorpo-rated(ESI) of Toronto, Ontario. Development ofthe concept started in 1996 and delivery of thecurrent system was completed in February 2002.

2preliminary investigation of an area suspected to bemined to verify whether it is mined and to better delimitthe area that contains mines

3detection sweeps following a clearance operation toconfirm the quality of the work

2. System Concept

The ARS relies upon a high resolution measure-ment of the terrain over which the detector is tobe passed to allow the computation of a trajectoryfor the detector payload. The terrain measure-ment is achieved by a scanning laser rangefindercoupled with several ultrasonic distance measure-ment units. The ultrasonic units are used to pro-vide for obstacle detection in directions that thelaser doesn’t scan and to provide height measure-ments in instances when the LRF fails to generatea measurement4. Figure 1 shows the configura-tion of the sensors. The laser range finder mea-sures the distance from the sensors to the groundthroughout an arc perpendicular to the motionvector of the arm. The spatial resolution of therange measurements varies due to the geometryof the sensors, but is nominally on the order of10 mm. Two ultrasound sensors are placed on ei-ther side of the main sensor head providing singlepoint distance measurements in the line of the de-tector head motion. These measurements do nothave similar spatial resolution but partially com-pensate for dropouts in the laser measurements.Two additional ultrasound sensors are used forobstacle detection for the laser arm itself.

LASER

ULTRASONIC SENSORS

LRF ARM

Figure 1: Terrain Sensor Configuration

A partial representation of the kinematics ofthe ARS can be seen in Fig. 2. The design in-cludes two “arms”; one carrying the detector pay-load and the second (the LRF arm) mounting theterrain sensors.

The detector arm moves the detector payloadin a circular arc about the shoulder joint (rotationq1) while the height of the detector is governed byrotation q2. The detector arm can be rotated (q3)to roll the detector and a linear actuator mountedon the arm (q4) allows the detector payload to betilted in the “pitch” orientation. The upper arm

4primarily when the terrain surface generates a specularreturn as can be the case with standing water

Turret

Shou

lder

Detector

Detector Arm

(LRF Pan)(LRF Tilt)

(Wrist Roll)

LRF Arm

(Wrist Tilt)

Figure 2: Robotic Scanner Kinematics

rotates with the shoulder; however this arm hasan additional degree of freedom to allow the ter-rain sensors to “lead” the detector and to gener-ate terrain measurements for the full width of thedetector sweep.

This representation omits an additional degreeof freedom which allows the shoulder to be trans-lated in the longitudinal (x) direction. The entireshoulder / arm assembly is mounted on a platformcoupled to a linear actuator that allows the shoul-der to be moved forward by up to 0.5 m. This de-sign allows the ARS to scan an area forward of itsmount point that is approximately 2 m wide and0.5 m deep (although highly non-rectangular).

In operation, the ARS is mounted on a vehi-cle. The vehicle and scanning system are thenremotely operated by the system operator. Theoperator manoeuvres the vehicle to an area of in-terest and initiates a scan sequence. The vehicleremains stationary during a scan by the ARS andthe data is telemetered back to the control stationfor display and interpretation.

A typical scan sequence would start with thedetector head in a ready position, raised well offthe ground. The laser scanner would then mapa small area at one corner of the available scanarea. The detector head is then lowered into thisarea and the laser scanning arm moved to lead thedetector motion as the detector is moved to theopposite lateral extreme. The shoulder positionis moved forward and the laser scanner moved tothe opposite side of the detector head to againlead the detector motion. In the current softwareimplementation, the computation of the detectortrajectory is a real time process using only theheight data from the area directly in front of thedetector. Data from previous scans is not uti-lized. This avoids any issue of data misregistra-tion should the orientation of the ARS change.

It should be noted that while the ARS includes

the ability to operate in a terrain following mode,it is equally possible to sweep the detector headin a plane, should that be desirable for interpre-tation of the detector data. The plane can becomputed based on the terrain measurements toprovide the closest approach to a point of interest,or to maintain a nominal height above the localground surface.

The ARS is designed to carry a payload of upto 5 Kg. For the purposes of the initial devel-opment and testing a Minelabs F1A4 detector isintegrated. This detector is capable of generat-ing a serial data stream representing its response,sampled with a period of 16 ms5.

The control station receives time aligned detec-tor data and joint position data. This is usedto generate a spatial display of detector responseversus position in the scan area. Combined withhigh resolution navigation data and motion esti-mates from the host vehicle, the detection datafrom multiple scans can be combined to providea map of detector response over a broad area.

3. Trial Plan

The trials plan developed for the ARS system en-compassed evaluating the performance of the armin both a laboratory environment and in a limitedrange of outdoor environments. The principal ob-jectives of the trials included:

• the ability of the laser scanner map buildingprogram to measure terrain features;

• the ability of the arm to control the detectorhead over terrain at realistic coverage rates;and,

• the ability of the arm control system to gen-erate position information for point targetsdetected by the metal detector sensor.

The laboratory tests also provide an opportu-nity to verify all aspects of the integration be-tween the ARS itself, the control system of thehost vehicle and the control station. Integrationissues include power consumption, time synchro-nization capability, command protocols, data pro-tocols, and error reporting.

As an initial step to validating the utility of theARS design, a variety of terrain types are used fortrials including:

• sand

• broken or damaged pavement;

5due to limitations of the embedded controller used inthe ARS the Minelab data is actually telemetered andrecorded at a period of 33 ms

DoF Range Referenced toq01 +78 to -88◦ centrelineq02 5 to 35◦ below horizontalq03 +55 to -55◦ horizontalq04 +12 to -13◦ horizontalq05 +25 to -15◦ lead / trail main armq07 0 to 500 mm stowed position

Table 1: System Capabilities

• gravel roads, both graded sections and sec-tions subject to potholes, wheel ruts andwashboard;

• coarsely mown grass or vegetated surfaces in-cluding examples with tracks, ruts and ex-posed rocks;

• surfaces with significant slopes or undula-tions; and,

• surfaces including obstacles such as markingstakes and trees that encroach on the scanarea.

4. Trial Status / Results

Trials of the ARS system are currently underway.Much of the laboratory testing has been com-pleted and preliminary results are available. Ini-tial laboratory testing was completed in a sand pitwithin a greenhouse complex. While much of thegreenhouse structure is metal, it is sufficiently faraway from the soil pits to provide an essentiallymetal free environment for the trials. Additionaltrials have been conducted in a conventional labo-ratory environment where the entire ARS assem-bly was raised and the detector was swept over anonmetallic surface with wooden obstacles addedas required.

The system capabilies are summarized in Ta-ble . For most of the trials performed to date thescan arc movement limits were limited to 40 de-grees either side of centre. The nominal terrainoffset used was 50 mm. The scan speed was 0.5m/s6. The longitudinal position of the sweep arcwas advanced by 30 to 50 mm between sweeps.For trials aimed at assessing the terrain followingperformance the detector position was advanced(in the x axis) at the end of a sweep in either di-rection, whereas, for trials aimed at collecting de-tector data the sensor head position was advancedfollowing a sweep sequence in both directions.

Trials of the terrain following behaviour havebeen completed over a variety of surface profiles.

6this is the maximum sweep speed achieved during ascan. Acceleration profiles reduce the speed near the endof the sweep arcs

Figure 3: Example of Terrain Following Test

The initial trials were conducted in a sand soilwith sand removed to form various surface pro-files, including ruts and potholes. A photographof a typical trial surface is included as Figure 3.The trials were generally successful with the de-tector able to follow most profiles attempted.While the performance of the ARS is currentlyadequate to support trials, some limitations ofthe current implementation have been identified.These primarily relate to the control algorithmsdriving the trajectory of the detector head. Thecurrent control strategy doesn’t explicitly use theterrain measurements to form a terrain elevationmap, but rather, uses the terrain data as a setpoint input to a relatively simple closed loop con-trol strategy7. This limits the types of terrainthat can be accommodated to those that are es-sentially “smoothly varying”. Other issues withthe current implementation relate to difficultiesfollowing terrain excursions near the edge of thescan path. This is due to initialization artifactsas the scan commences. None of these limitationsis fundamental to the concept, but would requiresignificant revision of the control software to fullyexploit the mechanical capabilities of the system.

5. Example Detection Results

Detection data was collected for a variety of tar-gets using the ARS. Tests included scans overboth flat surfaces – in which the detector headwas nominally in the same horizontal plane at alltimes – and over terrain features where the detec-tor head followed the terrain profile.

Example plots of detector response versus de-tector position are included as Figures 4 and 5as an intensity image. Both of these plots arederived from the same target8, in the same po-

7in reality even the current control implementation isquite complex due to extensions required to accommodatethe spatial extent of the detector head

8the target for these plots is a “clutter” example – a

sition(equivalent to “pixel” 14,14 in the images).Figure 4 is derived from data collected as the de-tector head moves from right to left across thetarget and Figure 5 represents the response asthe detector moves from left to right9. The differ-ences between the two plots result from the timedomain behaviour of the Minelabs F1A4. The de-tector has significant delay in the signal process-ing chain between the time the target enters theresponse zone and an output response. Furtherthe detector exhibits an even longer decay timefor the response. Human operators successfullycompensate for, and even exploit, this behaviourin their use of the detector; however, it doesn’tlend itself to straightforward spatial data inter-pretation. Despite the significant differences, it isstill possible to localize the target from the combi-nation of the two spatial data sets. The target lo-cation (for a point target) is essentially the pointof symmetry between the two images. Displaytechniques to allow easy operator interpretationof the spatial result are being explored, but haveyet to be finalized.

For comparison, Figures 6 and 7 show data col-lected with a larger target (a crushed aluminumpop can in this instance – centred at “pixel”14,14). Localization is still possible, but the ex-tent of the signature is so large it extends some-what beyond the scan area of the ARS (at a singleposition). Other extended targets, such as largerpieces of scrap, loops of wire, metal anti-vehiclemines or larger EOD items can exhibit signifi-cantly larger spatial signatures. This emphasizesthe requirement to being able to register data col-lected during multiple sweeps.

6. Future Work

Many aspects of the trial program remain to becompleted. Controlled laboratory trials will beconducted against a variety of targets to collect amore complete data set to validate data visualiza-tion methods supporting target localization. Thisdata set will also be exploited to develop conceptsand techniques for interpreting data to infer tar-get characteristics such as size, burial depth andmaterial. Data interpretation may ultimately beautomated, or it may remain as cues and guidanceprovided to an operator for manual interpretation.

The majority of the effort remaining to com-plete trials for the ARS is related to integratingthe system onto a remotely operated vehicle andcoupling the ARS control into the vehicle controland navigation systems. Mechanical integration

tab from an aluminum soft drink can9the figures show the data with increasing longitudinal

dimension down in the page. The ARS “shoulder” is atthe top of the image; hence, “left and right” are reversedin the image

5 10 15 20 25

2

4

6

8

10

12

14

16

18

20

Figure 4: Detector Output Image (tab – LeftScan)

5 10 15 20 25

2

4

6

8

10

12

14

16

18

20

Figure 5: Detector Output Image (tab – RightScan)

is complete and the ARS is shown on one of ourresearch vehicles in Figure 8. Developing exten-sions to our existing vehicle control architectureand operator control station to support controlof the ARS and display of the data is underway;however, the data display has not been validatedas yet. The significant unknown in this regardis whether it will be possible to merge detectiondata from several passes of the ARS to form aspatial detection map over a larger area. This isrequired to localize targets near, or at the edge of,any given scan. While this is conceptually simple,and has been done in other systems for lower reso-lution data, the high spatial resolution of the datacollected by the ARS will expose errors in positionestimation to a great degree.

Once integration has been completed, trials willbe conducted to investigate how best to exploitthis class of system within a mine action environ-ment and to fully define the potential utility ofthe system.

5 10 15 20 25

2

4

6

8

10

12

14

16

18

20

Figure 6: Detector Output Image (can – LeftScan)

5 10 15 20 25

2

4

6

8

10

12

14

16

18

20

Figure 7: Detector Output Image (can – RightScan)

Figure 8: ARS Mounted on “Scout” RemotelyOperated Vehicle

7. References

[1] J.E.McFee, V.Aitken, R.Chesney, Y.Das,and K.Russell, “A Multisensor, Vehicle-mounted, Teleoperated Mine Detector WithData Fusion,” in Detection and Remedia-

tion Technologies for Mines and Mine-likeTargets III, A.C.Dubey, J.F.Harvey, andJ.T.Broach, eds, Proc. SPIE Vol. 3392, pp.1082-1093, (Orlando, FL, USA), 13-17 April1998.

[2] T.Hanshaw and D.M.Reidy, “OperationalStandoff Mine Detection: Its Technology andApplication,” in Detection and RemediationTechnologies for Mines and Mine-like Tar-gets II, A.C.Dubey, R.L.Barnard, eds, Proc.SPIE Vol. 3079, pp. 432-442, (Orlando, FL,USA), 21-24 April 1997.

[3] T.J.Gorman, “Analysis of Sensor Integra-tion of the Integrated Ground Mobile MineDetection Testbed (IG-MMDT),” in De-tection and Remediation Technologies forMines and Mine-like Targets II, A.C.Dubey,R.L.Barnard, eds, Proc. SPIE Vol. 3079,pp. 443-451, (Orlando, FL, USA), 21-24April 1997.

[4] Y. Das, K.Russell, N. Kircanski and A.A.Goldenberg, “An articulated robotic scannerfor mine detection - a novel approach to vehi-cle mounted systems”, in Detection and Re-mediation Technologies for Mines and Mine-like Targets IV, A.C.Dubey, J.F.Harvey,J.T.Broach,and R.E. Dugan eds, Proc. SPIEVol. 3710, pp. 887-894, (Orlando, FL,USA), April 1999.

[5] Al Carruthers, John McFee, Denis Bergeron,Yoga Das, Robert Chesney and Kevin Rus-sell,“Scoping Study for Humanitarian Demi-ning Technlogies,” Suffield Technical ReportTR 1999-121, DRDC Suffield, Medicine Hat,AB, Canada, September 1999.

AUTOMATED MINE DETECTION ALGORITHMS FOR SOILS WITH MINERAL CONTENT

Olga Duran, Kaspar Althoefer and Lakmal D. Seneviratne

Department of Mechanical Engineering, King's College London

The Strand, London WC2R 2LS, UK, Tel: +44 20 7848 2431 {olga.duran, k.althoefer, lakmal.seneviratne}@kcl.ac.uk

Abstract - The standard mine detection approach is to employ a hand held induction-based mine detector that is moved over the ground. A major problem in achieving a close to 100% accuracy is the presence of ambient signals caused by mineralised soil or shrapnel leading to false alarms. This paper investigates changes in the induced field due to metal components such as those contained in mines, in order to develop automated mine detection algorithms for soils with mineral content. Algorithms for automated mine detection in the presence of mineralised soil are studied. The signal processing, feature extraction and intelligent interpretation software are introduced. Keywords: Mine detection, Automation, Artificial Neural Networks, Metal Detector. Introduction More than 100 million mines have been laid in the world, presenting a latent danger for civilians and soldiers [1]. The current mine detection strategy in to manually scan areas that are suspected to contain mines. With mine detection it is vitally important to have near 100% detection accuracy. Positive false alarms extend the inspection time considerably and negative false alarms put the operator at risk. A major problem in achieving high accuracy is having automated mine detection strategies in the presence of ambient signals caused by mineralised soil or shrapnel. A detection system, which can reliably distinguish between mines and soil with metal content, is required to keep the inspection time low and, at the same time, increase the safety of the operator. Currently, induction-based metal detectors are the standard sensor used for mine detection. Investigations carried out recently suggest the use of a second sensing method based on vision and sensor fusion algorithms for data interpretation and decision [1-2-3]. Ground Penetrating radar (GPR) is often chosen for this application [1-2]. However the economic restrictions make very difficult the use of such techniques since they can increase the inspection price considerably. Inspection solutions

using metal detector-based sensors combined with intelligent data analysis and classification and providing high accuracy are required. In this paper, standard mine detection data processing capabilities are studied. In Section 2 a description of sensor data is presented. Section 3 deals with data processing. Preliminary results using Artificial Neural Networks (ANN) are presented in Section 4. Finally conclusions are given in Section 5. Sensor description Metal detectors are usually based on pulse induction techniques. A pulse of current is sent repeatedly through a coil. This current commonly raises relatively slowly to a certain level. However, at the end of the pulse, the current is switched off very rapidly inducing a very large voltage spike or back e.m.f across the coil. No current flows through the coil after this transient. When a conductor target is present, as soon as the pulse stops, eddy currents are induced. These eddy currents always flow in such a direction as to re-create the magnetic field that has just disappeared. Eddy currents will generate a magnetic field around the target. Once applied and sampled, this signal is used to indicate the presence of metal [4]. Figures 1 and 2 show the resulting signals in presence of a metallic target and in ferrous soil respectively.

Figure 1: Sample signal across a metallic target.

Figure 2: Sample signal across ferrous soil. The peak does not correspond to a target but is

produced by the conductive soil.

TABLE 1 : DATA CHARACTERISTICS

Targets with different quantities of metal

Max. Amplitude

Min. Amplitude

Width of Peak

Target1 (experiment a) 9925.25 -9900.25 6.63

Target1 (experiment b) 9925.00 -9997.25 7.50

Target2 (experiment a) 151.25 -96.50 -

Target2 (experiment b) 5082.25 -602.50 14.25

Target2 (experiment c) 178.00 -155.00 -

Target2 (experiment d) 2408.00 -657.75 24.25

Target3 (experiment a) 9786.00 -7960.25 11.75

Target3 (experiment b) 461.50 5.50 51.25

Target3 (experiment c) 97.00 -90.50 -

Target3 (experiment d) 22.25 -48.75 -

Target4 (experiment a) 9999.00 -432.00 25.50

Target4 (experiment b) 9999.00 -10000.00 31.00

Target4 (experiment c) 9998.75 -9936.25 22.50

Target4 (experiment d) 9999.00 -10000.00 18.75

Ferrous Soil (exp. a) 9703.50 -4153.00 11.50

Ferrous Soil (exp. b) 4951.50 -1725.75 7.75

Ferrous Soil (exp. c) 1291.75 -1403.00 17.75

Ferrous Soil (exp. d) 167.00 -84.50 22.75

Ferrous Soil (exp. e) 34.50 -56.50 14.00

Ferrous Soil (exp. f) 54.50 -57.00 -

Ferrous Soil (exp. g) 990.00 -907.00 11.25

Ferrous Soil (exp. h) 510.75 -401.00 11.75

3. Test procedure and feature extraction The experiments were conducted using the induction-based metal detector MD8 Bravo, which is manufactured by the London-based company Guartel Ltd. During the experiments, the hand-held detector was moved over the target area producing signals like those shown in Figures 1 and 2. Signals were recorded for different targets in ferrous and in non-ferrous soil, as well as for ferrous soil containing no target. Table 1 describes the characteristics of the signals in presence of different metallic targets [7]. The targets are representing the metal contents in landmines. In these experiments, the induction signals in response to targets of different sizes were investigated. It is observed that the changes in signal amplitude are mainly related to the quantity of metal in the targets. Targets 1 to 4 are metallic targets with different quantities of metal, while the ferrous soil does not contain any metallic target. The higher the quantity of metal in a traget is, the higher is the measured amplitude. Figure 3 represents a map where maximum and minimum amplitudes are plotted against each other. It is noted that 3 zones can be identified. Zone A contained data form targets with high quantity of metal (Target1 and Target4). Zone B represents metallic targets with different quantities of metal. Zone C represents ferrous soil in the absence of a target. Except for some samples corresponding to targets with a low quantity of metal that go into Zone C, the amplitudes tend to give a good estimate. However, fixed thresholds would not give accurate results, as there is not a 100% correlation between the zones in the map and the targets. The ferrous soils are always identified. However, targets with low amount of metal are wrongly identified as ferrous soil with no target in 31% of the cases. Artificial Neural Networks are tried to give a more accurate correlation of the Amplitudes versus targets in Section 5. 4. Sensor data processing It is observed from table 1 and figures 1, 2 and 3 that there is a relationship between the symmetry of the signal and the presence of a target. Negative and positive peaks tend to have symmetrical shapes in the presence of a metallic target while asymmetries appear when ferrous soil exists. In order to emphasise this, the ratios of the positive and negative peaks in the signal are calculated (see figure 4).

Figure 3: Map with Maximum Amplitude plotted against Minimum Amplitude. Three Zones can be identified in the map, corresponding to highly metallic target (A), metallic target with a lower metal concentration (B) and ferrous soil (C) respectively.

Figure 4: Top: Signal corresponding to a metallic target (the positive and the negative peaks are virtually symmetrical to each other); Bottom: Signal corresponding to ferrous soil (the positive peak is much higher than the negative one).

Figure 5: Map with Positive Peak Ratio plotted against Negative Peak Ratio. The data corresponding to ferrous soil (A) can be clearly identified. Nearly all signals referring to soil are correctly identified using the threshold method.

Figure 5 summarises the results obtained with this technique. In our experiments, this technique showed the data corresponding to ferrous soil could be identified from the one corresponding to metallic targets with high accuracy. Only around 5% of the signals related to ferrous soil was missed.

5. Classification using a neural network The aim of the neural network (NN) is to classify the different targets from the signal produced by the mine detector. This work focuses on the identifi-cation and classification of targets with different metal content and soil qualities. Sets of pre-processed information representing the data generated by the detector are fed into the network. The main aim of the pre-processing stage is to remove measurement noise from the training data, that could otherwise enter the NN algorithm and interfere in the learning process and probably complicate the classification task. Here the amplitudes are used as the pre-processed inputs. 5.1. Neural Network input From the analysed data presented in Section 3, it can be concluded that a simple approach could use the maximum and minimum amplitudes as input for the NN. Figure 3 shows that there is a relationship between the amplitudes of the signal and the quantity of metal in the targets. In the “Max Amplitude against Min Amplitude” map, a number of zones can be identified.

5.2. Network structure The network used in this work is a multi-layer perceptron (MLP). It consists of three layers composed of neurons. These neurons are arranged in such a way that each of them has a weighted connection coming from every neuron in the previous layer. Each neuron performs a summation of all its inputs and passes the value through a non-linear function before sending it to the output. In this study, the number of nodes per layer is chosen using a trial and error procedure. The ANN consists of an input layer, an output layer and one hidden layer. The ANN is fed with sets maximum and minimum amplitude values as shown in section 3. A scaled conjugate gradient (SCG) training algorithm is used in this study. SCG is based on backpropagation, widely used for solving classification problems [5-6]. The training rule is given by:

kk P⋅+=+ αk1k W W (1)

where k is the presentation number, Wk is the current weight, ∇k is the learning rate and Pk is the weight search direction. The algorithm starts by searching the steepest descent direction, 0P (negative of the

gradient 0G ):

00 GP −= (2)

Then, the procedure determines the next weight search direction combining the previous search direction and the new steepest descent direction (current Gradient G ):

1kk G- P −⋅+= kk Pβ (3)

The constant kβ is computed as follows:

21

2

−

=k

kk

G

Gβ (4)

Where • is the norm of a matrix, and kβ is the ratio of the squared norm of the current gradient ( kG ) to the norm squared of the previous gradient ( 1−kG ).

5.3. ANN design In this investigation, a binary classification between target and non-target in the presence of ferrous soil is tried. The best performance was achieved with a multi-layer perceptron with three layers: the first layer with 2 input neurons receiving maximum and minimum amplitudes as input, a second hidden layer with 50 neurons and one output neuron. The output neuron values '1' or '0' correspond to object and soil respectively. The data used during training consists of 50 data files, corresponding to both object and soil data. During training, the error is brought down to 1e-10 after 12000 epochs.

Figure 6: The sum-squared error of the ANN during training (binary classification).

Figure 7: Generalisation behaviour (binary classification).

Figure 7 shows the classification results using unseen data. It shows that the ANN can distinguish between the two targets with a 90% success rate.

6. Conclusions This paper describes a neural network based approach for automated mine detection. The focus of this work is on providing intelligent algorithms to process and classify data acquired from a commercial induction-based mine detector. Specifically the detection of metallic targets in the presence of ferrous, conductive soils was investigated. Although these are initial results only, this paper has shown that the chosen neural network based approach is promising and capable of processing the data obtained from the sensor and automating the target classification process. The neural network-based approach was also compared to an approach using simple thresholds. The presented network-based approach showed to be superior to the threshold-based approach with a success rate increase of 21%. Processing stages based on ratio calculations have been applied to the mine detector data and encouraging results have been found. Future work will deal with the selection of signal features such as the ratios, the combination of such data with ANN, and with the investigation of multiple classification problems in mine detection. 7. Acknowledgements The authors would like to thank Ms Maggie Wu for providing sensor data acquired during her studies. The authors are thankful to Guartel Ltd for providing the mine detector MD8 Bravo and monitoring software. 8. References [1] Gader P.D., Keller J.M, Nelson, B.N.”

Recognition Technology for the Detection of Buried Land Mines”, IEEE Transactions on Fuzzy Systems, Vol. 9, No. 1, pp 31-43, 2001.

[2] Gader, P.D. Nelson, B.N. Frigui, H. Vaillette, G. Keller , J.M. “Fuzzy logic detection of landmines with ground penetrating radar” Signal Processing 80, pp 1069-1084, 2000.

[3] Filippidis, A. Jain, L. C. and Martin, N. “Multisensor Data Fusion for Surface Land-Mine Detection”, IEEE Transactions on Systems, Man, and Cybernetics — part C: applications and reviews, vol. 30, no. 1, pp 145-150, 2000

[4] “About Pulse Induction” Available on –line : http://www.protovale.co.uk/tech/abtpi.html

[5] Haykin, S. Neural Networks. A Comprehensive Foundation. Macmillan College Publishing Company, NY, 1994.

[6] Moller, M. F., "A scaled conjugate gradient algorithm for fast supervised learning," Neural Networks, Vol. 6, pp 525-533, 1993.

[7] Wu, M. “Automated Mine Detection”, MSc Project Report, King’s College London, 2002.

AUGMENTED TELE-OPERATION

Robert ChesneyCanadian Centre for Mine Action Technologies

Abstract

The Canadian Centre for Mine Action Technolo-gies1 (CCMAT) was established in 1998 with thegoal of developing technology for humanitariandemining. One component of the research pro-gram is the development of an effective, generic,tele-operation kit for demining applications.

Co-located with the Suffield laboratory of De-fence R&D Canada (DRDC Suffield), CCMATdraws on extensive experience in the develop-ment of tele-operated systems for military ap-plications, including remotely operated mine de-tection systems. Combining this experience inthe development of tele-operation systems with arequirements analysis for mechanized equipmentin humanitarian demining, a systems concept forgeneric control systems was developed and imple-mented for trials and demonstrations.

A critical aspect of the systems concept is to ex-pand or augment conventional tele-operation con-cepts with targeted automation features. Strikinga balance between system capabilities and systemcomplexity, the system is designed to maximize ef-ficiency and reduce operator workload within anaffordable and maintainable system architecture.

The paper describes the requirements defi-nition process, the requirements identified, thesystem architecture and its current state ofdevelopment, implementation and testing.

Keywords: Demining, humanitarian, Tele-operation, CCMAT, remote control

1. Background

Humanitarian demining is an onerous task, char-acterized by extremely slow, labour intensive pro-cedures with significant risk to the demining staff.The extent of the areas that are mined, or thatare potentially mined, is large enough that currentclearance methods cannot resolve the problem inany realistic timeframe. Mechanically aided clear-ance methods, where machines assist in clearingvegetation prior to mine clearance, or in the de-tection and destruction (or removal) of mines hasgreat potential to speed the process to the ex-tent that large area clearance becomes possible.

1Box 4000 Medicine Hat, Alberta, CANADA

Unfortunately, most mechanical methods demon-strated to date have a significant potential for ini-tiating the mines that they are intended to re-move, either as a consequence, or as a side effectof the design. This requires that the operator beprotected by a heavily armoured cab, or that thesystem be operated by remote control. Armourprotection adds weight to the system and is diffi-cult to adequately qualify for full protection un-der all circumstances; often leading to the use ofremote control.

Once a system has been prepared for remotecontrol, it becomes possible to insert relativelylow cost technology to add automation, or roboticfunctionality, to the system. While a competentoperator can perform most required functions byremote control of the system’s actuators, auto-mated control of some, or all, of the system func-tions can result in faster operation or fewer errors.Further, an automated system allows integrationof data from position measurement systems anddetector systems that is difficult for a human op-erator to assimilate.

The staff assigned to support CCMAT in thiswork had extensive experience in the develop-ment of tele-operation systems for military ap-plications, including remotely operated landminedetection systems [1], however it was recognizedthat the humanitarian demining problem wouldraise different requirements than the military role.Hence, while relying on previous expertise the de-velopment of the concept and the hardware imple-mentation for augmented tele-operation for hu-manitarian demining was based on a thoroughrequirements analysis that examined the opera-tional context and constraints associated with thehumanitarian mine clearance efforts.

As a part of the requirements analysis forthis development certain observations were madeabout the typical operational environment. Theseobservations provide motivation for many aspectsof the implementation design. An overview ofsome elements of the operational context is in-cluded below.

1.1 Environment / Terrain

Mine clearance is an issue in many parts of theworld ranging from temperate European climatesthrough tropical and desert climates. Equipment

operation would be required under a broad rangeof environmental conditions. However, operationin weather extremes, such as high rainfall rates orobscuring fog is unlikely.

The terrain the systems would have to op-erate over would be similarly varied, rangingfrom desert sand, through rocky soils to tropi-cal swamps. Most operations would be expectedto focus on trafficable routes or arable land, butslopes encountered may be significant and difficultfor mechanized vehicles to traffic. Soil penetra-tion resistance may range from easy to extremelydifficult depending on soil type and moisture con-ditions. Some operations would be in built up ar-eas, operating near buildings, on or near surfacedroads. Heavy damage to road surfaces and sub-stantial rubble from collapsed or damaged build-ing would often be encountered.

Vegetative cover may range from none, to verysubstantial tropical growth. In keeping with thelikelihood that many operations will occur onarable land, significant growth would have to beassumed as the norm.

It is assumed that most remote control opera-tions could be conducted with direct line of sightto the clearance equipment.

1.2 Crew Expertise

Crews assigned to mine clearance can not be ex-pected to have a high degree of technical exper-tise relating to robotics, although they may be ex-tremely competent in other areas. A substantialcomponent of the crew will be indigenous to thearea and may have little, if any, technical back-ground. Crew members brought in from outsidethe area are more likely to be selected for their ex-perience and expertise on mine systems and mineneutralization than for their computer or electron-ics skills. System specific training will likely belimited and staff turnover may be high and un-predictable.

1.3 Threat

The majority of the threat mines encountered areexpected to be anti-personnel mines, with a mix-ture of blast and fragmentation types. Some anti-vehicle mines will be encountered, again with amixture of types possible. The mines may havebeen emplaced for substantial periods of time andthe fuzing mechanisms may no longer function asoriginally designed due to corrosion or other en-vironmental effects. This is not meant to sug-gest that the fuzes would not function, ratherthat they will be less predictable than recentlyemplaced mines might be. It is expected that asignificant number of A/P mines will be set up

with trip wire activation, primarily connected tofragmentation mines.

Beyond the emplaced mine threat, unexplodedordnance in general, may be a threat. Thiswould include unexploded bombs, bomblets, mor-tar shells and artillery munitions. Some of thesemunitions may be shallow, however, they couldreadily be quite deep (beyond 2 metres for bombsand heavy artillery).

1.4 Logistic support available

Logistic support is expected to be essentially lim-ited to whatever the mine clearance team bringswith it, or is capable of providing. Many clearanceoperations will occur in remote areas, with diffi-cult supply routes. Local acquisition of equipmentand services beyond simple mechanical repair andlabour can not be assumed.

Very limited commercial communications capa-bility can be assumed. Technical support require-ments for the equipment will have to be addressedby on-site documentation, or whatever trainingcan be provided to the clearance team before de-ployment. Again, local expertise to support elec-tronics, or complex electro-mechanical systems, islikely to be limited.

Electromagnetic interference or frequency allo-cation issues would not often be a problem, how-ever, fully defined spectrum allocation and controlcan not be assumed.

1.5 Measures of success

The goal in any clearance operation is to removeall of the mines, so that any subsequent user ofthe area will not be injured. Clearance percentageis therefore the most apparent measure of success.In reality, no clearance method will provide 100% removal under all conditions, however, currentmanual methods (if rigidly adhered to) can comeclose. Mechanically aided clearance may not beable to routinely achieve similar results, depend-ing on the method used.

Other criteria for success will be the clearancerate and clearance cost. Manual clearance rateshave a broad range, due to the impact of the soilcharacteristics and the extent of metallic clutterencountered. Further, the rate will depend onthe number of people assigned and the methodsused. Hence, clearance cost (dollars per hectare),may be the best measure of success. Reducing thetime to clear a given area is the most likely bene-fit for mechanical systems. Depending on systemacquisition, deployment and operating costs, thismay result in lower clearance cost. The ultimatedecision on the cost-effectiveness of mechanicallyaided clearance, will be scenario dependent, but

increasing clearance rate per unit system (life cy-cle) cost will always be a goal.

2. Applications Considered

The potential applications for mechanical aids inhumanitarian demining are very broad. Not all ofthe potential applications would necessarily bene-fit from remote operation, however, many would.Examples of applications considered include:

• vegetation cutters;

• flails;

• ploughs;

• excavators;

• area search detectors; and,

• local search detectors.

The applications considered were not based onan evaluation of the relative merits of differenttechniques, rather they are presented in the con-text of an analysis of the types of control require-ments that a generic machine of that type wouldneed. The requirements are divided into five tech-nical areas as follows:

• driving control requirements;

• navigation requirements, including path log-ging;

• actuator control requirements;

• visualization requirements; and,

• data interpretation support requirements.

Driving control encompasses all of the basiccontrol and monitoring functions associated withthe platform that the mechanical aid is mountedon. Navigation requirements refer to any require-ments to monitor or control the path of the vehiclein a specific frame of reference, be that in absolute”map” coordinates, a relative coordinate systemshared by several systems (including detector sys-tems), or in regard to the previous path of the ve-hicle. Actuator control refers to the manipulationand monitoring of the mechanical aid to move itrelative to the prime mover, or to control its ac-tions in general. Visualization requirements referto any operator requirements to see the operationof the vehicle, or implement, that can’t be metby observing the vehicle from a distance. Lastly,data interpretation requirements include any re-quirement for data processing or graphical datadisplay.

3. Control System Functional Require-ments

The requirements analysis led to a set of func-tional requirements for a generic control systemthat could be used to operate a broad variety ofmechanized equipment. These requirements in-clude:

• control of the valves, relays and actuatorsto operate both the supporting vehicle andwhatever implements may be attached;

• sensing feedback measurements for some ofthe actuators controlled;

• estimating the motion of the supporting ve-hicle (relative navigation);

• estimating the position of the vehicle (andpossibly the ”endpoint” of the implement) inan absolute or geo-referenced frame;

• operator controls for the system;

• indirect view of the area the vehicle is movinginto or that the implement is operating in;

• detection processing and spatial registrationof detection results;

• recording vehicle track, detector coverage,and detection data results as appropriate;and,

• integrated test tools for self test and systemfault diagnosis;

Beyond the specific functional requirementsidentified, several additional goals were identifiedfor the development work. Paramount amongthese goals was a desire to make the system easyto use. In addition, the system had to be as re-liable as possible and simple to maintain in theevent of breakdown. Cost reduction was also agoal. To this end a configurable system that couldbe augmented for more complex applications wasidentified as a requirement.

4. System Concept

There were sufficient parallels in the capabilitiesrequired for humanitarian demining, and thosepreviously developed for military system, that adecision was made to exploit existing Canadianmilitary system designs as a basis for develop-ment. While the military systems were devel-oped for the control of more complex systems thanthose envisioned for humanitarian demining it ispossible to remove high level functions that aren’trequired.

The system design separates into two logicalblocks, the on-board control system and the op-erator control station. The on-board control sys-tem is based on the Ancaeus control system de-veloped by DRDC Suffield. This system has beenused in numerous projects including the landminedetection system described in Reference [1]. Asapplied to humanitarian demining it has a struc-ture as shown in Figure 1. The minimum ele-ments of the system are a data link2 and theon board control core module. The core mod-ule is a custom designed single board computerbased on a capable micro-controller. The boardincorporates significant amounts of analogue anddigital I/O, including several high current drivechannels. Four serial communication channels areavailable along with two controller area network(CAN) fieldbus connections. The serial and CANconnections allow the core module to augmented,where necessary, by navigation instruments or adedicated navigation processor where the systemdesign warrants. Where required by the visual-ization requirements, the core module can alsobe augmented by a video multiplexer and videotransmitter. The video multiplexer is used toallow several cameras to be mounted on the re-mote vehicle, with the view point selection con-trolled by the multiplexer (and ultimately the op-erator). The core module of the system was de-veloped jointly by CCMAT and Amtech Aeronau-tical3. The core module can be packaged into arelatively small box (12 by 18 by 24 cm)and theelectronics are cooled through case conduction; al-beit with small fans to circulate air within theenclosure. The configuration built for CCMATuses military connectors, however, less expensiveindustrial connectors can be readily substituted.

The operator control station chosen to supportthis work is a version of the Vehicle Control Sta-tion (VCS) that has been developed jointly be-tween Defence R&D Canada and CDL Systems4.The block diagram for the operator control sta-tion is shown in Figure 2.

The software is portable to many computingplatforms, but for the CCMAT implementation aPC based computer (running Linux) is used. The“core” elements of the control station are the PCand the data link transceiver. A generic controlpanel is used to provide the operator with con-ventional joystick and switch controls. The con-trol panel is implemented as generic input devicesthat are assigned to control functions as part of

2many commercial data links are available, depend-ing on the transmission speed required and the frequencybands available

3Amtech Aeronautical Ltd, Medicine Hat, Alberta,CANADA – www.amtech-group.com

4CDL Systems Ltd, Calgary, Alberta, CANADA –www.cdlsystems.com

Figure 1: On-board Control System Block Dia-gram

Figure 2: Operator Control Station Block Dia-gram

software configuration; hence, the same hardwarecan be used or a variety of vehicles with only achange in the switch labels.

As in the case of the on-board control sys-tem, the control station configuration can be aug-mented to support video display for visualization.The system also supports operation of a differen-tial GPS base station to provide correction datato a high accuracy on-board navigation system, ifrequired.

5. Augmented Tele-operation

As noted above, ease of use was felt to be criticalto improving the efficiency of mechanized equip-ment in whatever role that it is employed in. Thisgoal is realized not only through conventional hu-man factors design of the equipment and operatorcontrols, but also in the integration of automationinto aspects of system operation that are difficultfor the operator. Typically these are tasks wherethe displacement from the vehicle removes con-ventional feedback, or in cases where it is difficult

for the operator to visualize and react to abstractinformation.

Simple examples of this type of targeted au-tomation include automated speed and steeringcontrol for the vehicle. In these modes the on-board control system manages engine throttle andgear selection, to maintain whatever speed the op-erator selects. Steering is also controlled to keepthe vehicle on a straight path, unless the oper-ator commands a turn. This frees the operatorto focus on “mission related” tasks; these mightinclude keeping a vegetation cutter at an appro-priate height, or monitoring a detector system.

Examples of higher level automation that canbe readily integrated include automated area cov-erage and actuator endpoint control. In auto-mated area coverage, the operator merely definesthe area for the vehicle to work in (perhaps to cutvegetation or to conduct an area detection scan)and the system manages the vehicle path to en-sure that the area is methodically covered, withappropriate overlap. Examples of the screens usedto define the area coverage and the resulting vehi-cle path plan are included as Figures 3 and 4. Inmany instances the operator will define the areaby driving the vehicle around the desired workspace and then use the vehicle path recording asthe input to the area coverage task. Once the areacoverage function is engaged the control systemmanages the basic vehicle path and speed, withthe operator guiding the vehicle around obstaclesand focusing their attention on payload operation.

Figure 3: Area Coverage Boundary Definition

Automated endpoint control is simply a methodof reducing the task complexity of operations that

Figure 4: Area Coverage Vehicle Path Plan

use a manipulator, such as an excavator. Ratherthan having the operator control the motion ofthe excavator through controlling the individualactuators, the operator control the actuator in aCartesian space relative to the vehicle (or rela-tive to his visualization view point). The controlsystem then computes the desired individual ac-tuator motions to achieve the desired goal.Thisreduces the training requirements for some typesof manipulators and offsets the lack of feedbackthat the operator suffers from in a remotely oper-ated system.

6. Demonstration Platforms

CCMAT staff have recognized that theoreticalconcepts are a “hard sell” in the demining commu-nity, so a portion of the research effort is devotedto implementing the concepts in hardware to al-low demonstrations in realistic environments. Tothat end a series of demonstrator activities are un-derway that exploit the tele-operation system dis-cussed above. Two demonstrations are targeteddirectly at demonstrating the tele-operation sys-tem capabilities and a third exploits the systemto demonstrate a scanning detector concept [2].The first two demonstrators are based on a com-mon platform, a Melroe BobCat. This vehicle hasbeen extensively modified to remove all of the op-erator controls and to instrument all of the hy-draulic actuators5. This vehicle can be readilyswitched between a variety of tools.

In the first demonstration, the vehicle will be

5current stroke position of the actuator is measured

equipped with a vegetation cutter. The auto-mated area coverage capabilities of the controlsystem will be exploited in this instance to in-vestigate to what extent an operator can be moreeffective in this role with that class of operatoraid.

The second demonstration on this platform willuse an excavator attachment (also instrumented).The goal of this demonstration will be to compareand contrast operator effectiveness and ease of usein endpoint and conventional control modes.

Figure 5: Automated Scanner Demonstration Ve-hicle

The last demonstration investigates the use ofan automated scanning system for conventionalmine detectors. This system, when mounted onan unmanned vehicle (as pictured in Figure 5)canbe used to detect and localize targets at a dis-tance. This project exploits the tele-operationsystem to allow rapid integration of such a com-plex payload.

7. System Status

The development of this system is nearing com-pletion. The control station software is merelyminor interface changes to an existing product,and the underlying software is commercially avail-able6. The onboard control system core systemhas been prototyped and is fully functional atthis point in time. Minor design revisions willbe incorporated in a second revision of the boardby the development contractor7. Commercializa-tion of the revised version may follow should thedemonstrations warrant.

The demonstration platforms are both in thefinal stages of integration. Actual trials of thesystems are expected to commence in November2002.

6from CDL Systems, of Calgary, Alberta, Canada7Amtech Aeronautical of Medicine Hat, Alberta,

Canada

8. References

[1] J. E. McFee, V. A. Aitken, et al A Multi-sensor, Vehicle-mounted, Teleoperated MineDetector With Data Fusion Conference onDetection and Remediation Technologies forMines and Mine-like Targets III, Proc. SPIE,3392, Orlando, FL, USA, April, 1998

[2] Robert Chesney, Yogadhish Das, TerrainAdaptive Scanning of Conventional Mine De-tectors IARP Workshop on Robots for Hu-manitarian Demining(current proceedings),IARP, Vienna, Austria, November, 2002

A COMPARISON STUDY OF OPERATOR PERFORMANCE WITH THREE CONTROLLERS FOR A REMOTELY OPERATED VEHICLE

Marcus D. Penny, Samuel Cotter, Nick Beagley, Neil Smith, Kwok Wong

QinetiQ Ltd, Farnborough, GU14 0LX, UK

Abstract An optimum design of a Human-Machine Interface is paramount for efficient and accurate control of a Remote Controlled Vehicle (RCV) by an operator during Explosive Ordnance Disposal (EOD). This comparison experiment measured the time taken to complete a typical EOD task using novel low profile and high profile joystick controllers developed by QinetiQ Ltd and the in-service Wheelbarrow RCV hand controller. The controllers interfaced to a Virtual Reality (VR) environment and allowed 20 operators to control a virtual Wheelbarrow performing an EOD task. The results reveal an equivalent reduction in EOD task completion time for the high and low profile joystick controllers as compared to the in-service controller (F = 6.13, df = 2, 36, p < 0.01). Keywords : Virtual Reality, Explosive Ordnance Disposal and Human Machine Interface 1. Introduction Explosive Ordnance Disposal (EOD) involves inspection, identification and destroying of bombs in a controlled and safe manner. EOD is an extremely stressful and difficult task to complete with Remote Control Vehicles (RCV). The UK armed forces use a hand controller with four thumbwheels to control the in-service RCV called Wheelbarrow through an EOD task. Cameras are mounted on the RCV to provide video picture feedback to the operator. These pictures provide the primary feedback that allows the operator to view obstacles and the environment around the vehicle. For a typical EOD task, the operator has to manoeuvre the vehicle from a safe location to the unexploded ordnance and then manipulate actuators on the RCV to position weapons on the device. The RCV operator must balance the conflicting demands of urgency and the need for extreme accuracy. Mistakes or collisions during an EOD task can be time consuming and in some case cause mission failures. The RCV has to be operated from a safe location and out of line-of-sight due to the safety implications of EOD work. Therefore the design of the RCV’s hand controller and its ease of use is all-important to a successful mission.

A simple and intuitive layout of the RCV functions on the controller is a critical element in reducing the workload and stress of the operator during EOD. The existence of cheap touch screens and joystick technology has enabled the design of the Wheelbarrow hand controller to be revisited. As part of the Ministry of Defence (MoD) research programme, QinetiQ Ltd have developed two novel controller concepts for a comparison, each having intuitive functionality to the operators and incorporate good ergonomic design. The consoles are based upon a combination of dual axis isotonic joysticks, camera controls and a touch screen for menus. Touch screens can provide flexibility of design but Parsons [1] has shown that touch screens were not ideal for some military tasks. However, Sears and Schneiderman [2] have concluded that touchscreens are more intuitive interface for pointing tasks compared to a mouse device and provide the potential for lower error rates. The design of the controllers was guided by Defence Standard 00-25 “Human factors for designers of equipment.”[3] This standard provides specific guidance for the designers of military equipment to ensure that the design meets the requirements and capabilities for its intended users. However trials were necessary to determine whether the new controllers improved the operators performance for the EOD mission and to optimise their design. This paper describes a comparison experiment between two new controller concepts and the Wheelbarrow hand controller that is currently used by the UK armed forces for EOD. The hypothesis for this experiment is that RCV controller with a touch screen and joysticks provides a simpler user conceptual model than a controller with thumbwheels and switched modes. 2. Method The experiment was designed to compare the in-service Wheelbarrow hand controller (HC) against the low profile (LP) and the high profile (HP) joystick controllers developed by QinetiQ Ltd. In a trial plan balanced against effects of condition, twenty EOD operators from the UK armed forces completed a typical EOD mission within a Virtual Reality (VR) environment. The task was completed

three times with each of the three controller types. The controllers interfaced to a VR environment that allowed the operators to control a virtual RCV through an EOD task as shown in Figure 1.

Figure 1 Photograph of low profile joystick console used during the experiment The use of a VR environment provided a cost-effective method of trialling and ensured repeatable conditions throughout the experiment. Time to complete the EOD task was recorded for analysis in this paper. The frequency and types of collisions with each controller was also recorded but will be published in a subsequent paper. Questionnaires were used to capture user preferences and their suggestions for improvements to each controller type after each set of experimental runs. 2.1. Participants 20 volunteers from the UK armed forces were used in the comparison experiment. No additional payments above the participants’ standard pay were made. Their ages ranged between 22-38 years old and all were male and right-handed. The experience of the participants operating the in-service Wheelbarrow ranged from 2 months to 7 years. All had received training on the in-service hand controller during their service. Therefore, no additional training on the hand controller was required. Prior to the experiment, the participants were briefed regarding the tasks they would perform and asked to sign an informed consent form. 2.2. Experimental Hardware The experiment required the in-service hand controller (HC), the low profile (LP) joystick controller and high profile (HP) joystick controller to control the virtual RCV in a VR environment.

The in-service hand controller as shown schematically in Figure 2 is fitted with mode scroll and selection buttons, four thumbwheels and a 125mm (diagonal dimension) Liquid Crystal Display (LCD).

Figure 2 Schematic diagram of the in-service hand controller The operator scrolls and selects between ten modes that determine the function of the four thumbwheels. The LCD displays the selected mode and a graphical representation of the RCV’s configuration. This graphic allows the operator to appreciate the position of the actuators when the RCV is out of sight. There are four single axis thumbwheels provide proportional control of the Wheelbarrow tracks, actuators and camera functions. The thumbwheels are centre sprung, with a dead centre zone to ensure that their demand returns to zero when released. The side thumbwheels 1 and 4 move through ± 90° about their centre position to control an actuator and are manipulated by the operator’s index finger. The thumbwheels 2 and 3 on the upper surface of the controller move through ±45° about their centre position to control an actuator or drive motor and are manipulated by the operator’s thumb. The hand controller was designed to be held by both hands. The LP controller is a console design with a 525mm (diagonal dimension) LCD and two joystick clusters as shown in Figure 3. Each cluster has three dual-axis isotonic joysticks (centre sprung) that are manipulated with the thumb, index or middle finger. The displacement and diameter of the individual joystick met requirements in both Def Stan 00-25 [3] and NASA STD 3000 [4], with a diameter of 9mm and a maximum displacement of ± 20°. However, a shaft length of 15mm was used to ensure that no accidental activation of a joystick. The orientation of the individual joystick within a cluster allowed the operator to control any joystick without the need for the operator to glance away from the LCD.

LCD

170mmSelectbutton

Scrollbutton

1 4

2 3Thumbwheels

Front View Side View

Thumbwheels

20mm

Figure 3 Photograph of Low profile joystick console Figure 4 shows a schematic diagram of a low profile joystick cluster. The right joystick cluster controlled the RCV drive and camera pan & tilt functions. The left joystick cluster controlled the actuator functions. All joysticks gave proportional output values between 0 and 1, where ‘1’ represents the full demand. The full demand from the joystick ensures the maximum speed of the vehicle or actuator stipulated in the virtual RCV model/algorithm.

Figure 4 Schematic diagram of Low profile joystick controller The HP controller has the same console design as the low profile controller in terms of LCD and Human Computer Interface (HCI) except for the joystick controller. The controller HP as shown in Figure 5, incorporates two small isotonic dual-axis joysticks (centre sprung) with the same characteristics as described for the LP and one whole-hand isotonic dual-axis joystick. The two small joysticks are mounted inside a whole-hand dual axis (centre sprung) joystick. One joystick is mounted so that it can be controlled by the thumb and other by the middle finger. The whole-hand isotonic joystick had a shaft length of 60mm, maximum displacement angle ±20° and a displacement force of 11.5N. The

right joystick cluster controlled the RCV drive and camera pan & tilt functions. The left joystick cluster controlled the actuator functions. Figure 5 Schematic diagram of Right-handed High profile

joystick controller The LCD display comprising of a touchscreen for the low and high profile joystick controller’s was organised into four quarters as shown in Figure 3. The top two quarters displayed the camera pictures from the virtual RCV. The bottom right quarter of the screen contains the graphical representation of the RCV configuration. The remaining quarter was reserved for the touch screen buttons to select high, low or neutral gears and vehicle configurations. The simulation computer was a Silicon Graphics Inc. Onyx RE2 and it provided the RCV configuration data and camera feedback to the controllers and the LCD. The simulation computer emulates the in-service RCV by mimicking the two real world camera pictures. The views from the virtual RCV were configured to the same horizontal and vertical Field Of View (FOV) as the real world RCV cameras. For this comparison experiment the two camera pictures were displayed on a 525mm (diagonal) LCD for all three controllers. 2.3. Virtual Reality Environment A VR environment was used to ensure repeatable conditions for all the participants and making the virtual RCV and the EOD task a constant throughout the experiment. A virtual Wheelbarrow was driven from the three controllers through an EOD task. Figure 6 shows the VR database and the virtual RCV.

Camera Views

Low-profile joystick controllers

Graphicdisplay

Touch screen buttons

170mm

47mm

47mm 68mm

Side ViewPlan View

60mm

65mm

30mm

80mm

95mm

End View

30mm

105mm

Side View

40mm

Plan View

60mm

25mm

Dual-axisisotonic joysticks

Whole-hand dual-axisisotonic joystick

Whole-hand joystick gland

Figure 6 Picture of Virtual Reality environment and the RCV The VR environment developed for this experiment comprised of three main elements: a terrain database, a virtual RCV and behavioural model/algorithms that allow the virtual RCV to interact with the terrain database. The VR terrain database used for this experiment was an accurate representation of a real world location and the virtual RCV was based on Computer Aided Design drawings of the real Wheelbarrow. The behaviour model emulated the main Wheelbarrow characteristics; RCV mobility, camera vibration during motion and terrain following. The full functionality and control characteristics of the Wheelbarrow, including actuators’ speed and range of movement and RCV handling, was included to ensure an accurate behavioural model of a virtual RCV used in the experiment. As a validation exercise, real world data was collected from 6 operators completing an EOD task with the RCV hand controller and the low-profile joystick controller. The time data collected from an EOD task performed in the real world was used to enhance the model’s fidelity and gather real world performance characteristics prior to the VR experiment. Measuring the accuracy of a virtual system against real vehicles was the essential first stage in developing a credible virtual reality that emulates the real world. 2.4. Procedures Each participant underwent a training process, lasting approximately one and a half-hours prior to the experimental runs. The aim of the training was to teach the participants the full functionality of the low and high profile joystick controllers and provide hands-on experience in using all three controllers operating a virtual RCV in the VR environment. To conclude the training, the participants were required to successfully complete a training exercise that tested the same skills required in the experiment EOD task. On satisfactory completion of the training exercise, the participants were briefed and asked to complete the VR experiment’s EOD task three times with each controller (condition). To minimise possible learning effects each participant was

assigned a balanced order in which to carry out the trial runs, see Table 1.

Participant No.

First Condition

Second Condition

Third Condition

1 HC HP LP 2 HP LP HC 3 LP HC HP 4 HC LP HP 5 HP HC LP 6 LP HP HC 7 HC HP LP 8 HP LP HC 9 LP HC HP

10 HC LP HP 11 HP HC LP 12 LP HP HC 13 HC HP LP 14 HP LP HC 15 LP HC HP 16 HC LP HP 17 HP HC LP 18 LP HP HC 19 HC HP LP 20 HP LP HC

Table 1 Balanced Experimental Design1 The task in the experiment required the operator to manoeuvre the RCV through a defined EOD task (see Figure 7). The RCV was required to pass through three gaps, including the negotiation of a roadside kerb en route to the unexploded ordnance (target). Once the RCV had been driven into position the operator was required to manipulate the actuators in order to position the weapons correctly. The target was a small package placed within a rubbish bin. The end of the EOD task was reached when the weapons were in the firing position. The complete EOD task was divided into six stages; Stage 1- Drive RCV 80m along road from start point; Stage 2- Drive to the kerb through parked cars; Stage 3- Pass through the gate; Stage 4- Manoeuvre the RCV through the front garden; Stage 5- Move to the back garden via the garden side path; Stage 6- Position the weapon on the target device.

1 HC= In-service Wheelbarrow hand controller, HP= High profile joystick controller and LP= Low profile joystick controller.

Figure 7 Schematic diagram of EOD task performed in VR experiment The total times to complete all six stages were used to compare the three controllers. In addition to the use of objective measures, the participants were interviewed in order to gain an insight into their experience and impression of the controllers. 3. Results 3.1. Analysis technique The effect of controller type and to run order was tested by Analysis of Variance (ANOVA) to identify any significant spread of data. Newman-Keuls range and Bonferroni tests tested further significant differences between runs and controllers respectively. 3.2. Time to complete the EOD task The mean times taken to complete all six stages of the EOD task for each controller over each run are displayed in Figure 8. The differences between the conditions were analysed using the Newman-Keuls test and it showed a significant effect of run, irrespective of condition (F = 37.19, df = 2, 37, p < 0.001). Run 3 was significantly quicker than runs 2 and 1. The mean of run 2 was significantly quicker than run 1.

Figure 8 Comparison of mean total times as a function of number of runs for each controller

The results also show a significant effect of condition (F = 6.13, df = 2, 36, p < 0.01). The HC was significantly slower than the HP and LP for each comparative run and the overall mean time for each of the three runs. The mean total time taken to complete the course was found to be significantly less when the participants used the HP or LP. The fastest of all the mean total times was with the HP on the third run, however the overall fastest mean total time of all runs was using the LP controller. There was no significant difference in completion time between the HP and LP controllers. The run times showed significant (p<0.001) improvements for each additional repetition of a controller type. However as discussed in previous paper [5], the LP and HP controllers would continue to offer enhanced operator performance if adopted into service. 3.3. Subjective Data from Comparison Questionnaire Table 2 shows the ranking (1st, 2nd, 3rd) of the three controllers across the range of questioned features. The HP controller was ranked number 1 for all of the performance measures, i.e. smoothest, most instinctive and most responsive. It was generally perceived that the HP enabled the best performance and the participants generally thought that the worst performance was experienced when using the LP. However, the LP option was the preferred overall, whereas the mode of responses for the HC did not rank top for any of the questions. The only positive comments regarding the HC were that it was familiar (n=6) and portable (n=1). The negative comments concentrated on the scrolling aspect of the HC (n=6). The LP was considered easy and simple to use (n=4), had a good layout (n=5), instinctive to use and every thing was to hand (n=1). These positive comments were duplicated for the HP, however there was criticism of the joysticks being too bulky (n=2), whereas the LP was too ‘fiddlely’ (n=1).

The operators ranked the HP or LP controllers over the HC on every issue. The LP was ranked first choice controller for long periods of use and was the overall preferred option. The HP mirrored many of the positive aspects of the LP, and vice versa, i.e. the touch screen display and ability to multitask. However, both controllers were criticised on the fine detail of the joystick feel and sensitivity. There was also some confusion with regard to the allocation of function. These comments can be used to differentiate between the controllers.

House HouseBuilding

ROAD

12

34

56

Car CarPavement

Target at stage 6

Route taken

300

350

400

450

500

550

1 2 3Run number

Tim

e to

com

plet

e EO

D ta

sk (s

econ

ds)

Hand Controller

Low Profile joystickcontrollerHigh Profile joystickcontroller

Questions Mode ranking

for the In-service Hand

Controller (HC)

Mode ranking for the Low

Profile joystick

controller (LP)

Mode ranking for the High

Profile joystick

controller (HP)

Smoothest Control

3 2 1

Most Instinctive

3 3 1

Most Responsive

2 1 1

Perceived best performance

2 3 1

Preferred for long use

3 1 2

Overall preference

3 1 2

Table 2 Mode ranking of the three controllers The objective data shows that the HP and LP are faster than the HC controller, and the subjective data reveals that the participants also preferred the HP and LP to the HC. There was one question from which the answers contradicted this trend. When asked, “Which controller do you think you performed best with?” The answers did not correlate to the actual performance. Without knowing the times and accuracy of the runs, participants believed that their worst performance was experienced when using the LP, when in actual fact the LP was as fast, and in some cases, faster than the other two controllers. This illustrates the need to conduct objective measurements through comparison experiments. 4. Conclusion From the objective measurements taken, this experiment has shown that both the High and Low profile joystick controllers enabled a participant to complete the EOD task in a faster time compared to the in-service hand controller (HC). This could be due to only having one joystick for the RCV drive function on the consoles that does not rely on the hand co-ordination required for the in-service hand controller that uses two separate paddles for left and right tracks. In addition, the console joystick layout provides the operator with all actuator functions to hand with out the need menu scrolling as incorporated in the hand controller. The DEF STAN 00-25 part 13 [3] states that modes are acceptable when used for long periods of time (few minutes), the mode scrolling on the HC may have contributed to the slower completion times in comparison to the

controllers that have touch screen interaction style, with all functions immediately available. From the subjective measures taken, both the HP and LP were preferred over the in-service HC. As there were no differences between the HP and LP in the objective data, other factors must be considered when choosing between the two controller types. The Human-Machine-Interface for both the low and high profile joystick controllers are the same design with the exception of the joystick configuration. It is therefore unsurprising that the completion times were statistically similar. The user feedback on the new controllers was of a constructive nature. Suggestions for improvements were often offered alongside the negative aspects, indicating that the users saw potential in the new controllers. When the HC was criticised the participants needed to be prompted to supply a suggestion for improvement. This often resulted in the comment that the in-service HC should be replaced by one of the new controllers. 5. Acknowledgements The UK Defence Procurement Agency (DPA) funded this work. We wish to thank the members of the UK armed forces for their participation in the experiment. 6. References [1] M. B. Parsons, (1994). “Performance of manual c

ontrols within armoured fighting vehicles”, Contemporary Ergonomics, pp. 150-154.

[2] A. Sears and B. Schneiderman, (1991). “High precision touchscreens: design strategies and comparisons with a mouse”, International Journal of Man Machine Studies, vol. 34 (4), pp. 593-613.

[3] Ministry of Defence, (1987). “Human Factors for Designers of Equipment”, DEF STAN, vol. 00-25.

[4] National (1994). “Joystick design”, NASA STD, vol. 3000, section 9.3.3.4.2.

[5] M. Penny, S. Cotter, N. Smith and K. Wong (2001). “Development and validation of an Explosive Ordnance Disposal (EOD) Virtual Reality (VR) environment used for a Human-Machine Interface experiment”, Virtual Reality in Mechanical and Production Engineering Conference, 22-24th November 2001, Belgium.

Generating Distance-Invariant Object Representations by Subsampling of Images

Minh-Chinh Nguyen

Department of Computer Science Technical University Munich

Boltzmannstr. 3, D-85748 Garching Email: Nguyen@in.tum.de

Abstract

A novel approach for compensating perspective dis-tortion by a depth-dependent image transformation has been developed. Its usefulness for recognizing objects that may appear in unpredictable and widely varying distances has been demonstrated by imple-menting it on a real-time vision system and testing it in real-world experiments. The key point of the ap-proach is a subsampling of the image with the inter-val between subsampling points in each small region of the image being inversely proportional to the local depth. The method has proved its effectiveness in applications, and it is fast and simple. Moreover, by reducing the number of pixels to be processed by subsequent stages of the vision system it can contrib-ute to additional gains in processing speed.

Keywords : Distance Invariance, Object Recognition, Robot Vision.

1. Instruction

Object recognition is a key problem in the design and operation of autonomous vehicles and robots. The objects to be recognized by a mobile robot working in office environments include objects that might constitute a collision hazard, objects that should be manipulated in some way, or, generally speaking, all static or moving objects that could be relevant for the behavior of the robot.

Such objects may sometimes be close to the robot, at other times the same objects may be far away, and in some situations they must be recognized even while their distance is changing rapidly. The great varia-bility of distances in which external objects may be seen by a robot is one of the factors making object recognition difficult.

If vision is the sensing modality used for object rec-ognition distance affects the appearance of an object in the following ways:

• The size of the image of an object depends on the distance; if Z is the distance between an ob-ject and the camera the number of pixels belonging to the object’s image is proportional to Z-2, and the linear size of the image is propor-

Z-2, and the linear size of the image is propor-tional to Z-1.

• Similarly, the minimum size of a feature on the object that the vision system can resolve is pro-portional to Z2, or Z, respectively.

• If an object moves relative to the camera and parallel to the image plane the resulting dis-placement (measured in „pixels”) of its image relative to the image frame, and thus the result-ing motion blur, is proportional to Z-1.

• The contrast within the image of an object de-creases as Z increases; quantitatively this effect depends in a complicated way on the lighting and on the conditions of the atmosphere, but un-der typical conditions the effect is fairly pro-nounced if the distance exceeds a few tens of meters.

Together, these effects make it difficult for a vision system to determine, for instance, whether two im-ages, both showing an object, but in rather diffe rent distances, show the same object or not. It also makes it difficult to determine if distances, e.g., d1, between nearby doors and distances, d2, d3, between distant doors (Fig. 1) in an image are the same or different. Generally speaking, images of distant objects tend to look like small gray blobs without much internal structure, while images of near objects tend to be large and are often cluttered with details.

Figure 1: Perspective distortion affects to a corridor scene.

2. Nature’s Solution

For human beings these difficulties apparently do not exist. We can easily recognize an object when we see it, regardless of its distance. Normally, we are not even aware of the fact that, for instance, the size of an object=s image on our retina depends on the dis-tance. Rather, we always see objects in their `real@ size, regardless of their distance. This phenomenon has been studied by psychologists, and it appears that the human visual system performs a distance-dependent transformation on the elements of an im-age as one of the first steps of object recognition.

A prerequisite for performing such a transformation is knowledge of the distance from the object. It is still largely unknown how the human visual system can determine distance from a 2-D image, but obviously this ability exists. [Rock 1985] cites disparity, con-vergence, accommodation, motion parallax, and „im-aging factors” as possible clues to distance. The im-aging factors include perspective, shadows, occlu-sions, and the actual size of known objects. When we look at photographs we can often estimate the dis-tance between the camera and the photographed ob-jects; this indicates that the imaging factors alone may suffice for such an estimate.

If an image of a given size is projected on the human retina the perceived size of the corresponding object is directly proportional to the perceived distance (Emmert’s law) [Rock 1985]. Numerous observations support this law. For instance, when we see the moon high up in the sky it seems to be much smaller than when we see it near the horizon. The explanation is that the moon is isolated from other objects when it is high on the sky, allowing us to believe that it is not too far away; when the moon is seen near the horizon, however, we see that it is farther away than any other object.

Optical illusions also support Emmert’s law; Ponzo’s illusion (Fig. 2), named after the Ital-ian psychologist Mario Ponzo who discovered it, is an example. Al-though the two hori-zontal lines in Fig. 2 are of equal length the upper one is perceived as being longer. The most accepted explana-tion (proposed in the 19th century by Armand Thiéry) is that the two slanted lines are sub-consciously perceived as perspective images of two parallel lines, such as the edges of a floor, and that then the upper „object” is perceived to be farther

away than the lower one. Since both „objects” yield images of the same size on the retina the upper „ob-ject”, being farther away, must be larger.

Such optical illusions support the assumption that the human visual system estimates the distance from objects that are being seen and uses this information for modeling external objects with their supposedly true size, regardless of the distance from which they are being seen. In the sequel we will introduce an ap-proach by which a machine vision system may simi-larly transform distant-dependent image data into an approximately distance-invariant form in order to model physical objects regardless of the distance from which they are seen.

3. Technical Solution

The problem of modeling and compensating the ef-fects of perspective distortions in the image to the appearance of an object was initially addressed by [6] and [5]. For instance, to track and classify objects, especially motor cars, from within a car running, sometimes at high speed, on a highway [6] have used two separate models for their appearances, one for the near range and one for the far range. In the simi-lar manner, in the object manipulation domain differ-ent object models have been used by [5] to recognize a variety of differently shaped objects used for ma-nipulating objects by calibration-free visual manipu-lators.

In principle, such an approach raises the question which one of the models should be chosen in any specific situation. Fortunately, the model selection turned out to be not too difficult in this specific envi-ronment, as either one of the models could be used in an intermediate range of substantial extent.

Nevertheless, using distinct models for specific dis-tance ranges may not be the best possible solution. A novel approach that may be assumed to exist in a similar form in organic vision system [1], [7] is, therefore, investigated. Key points of this approach are:

• A transformation for compensating the effects of perspective distortion is performed on that section of the image which contains the object candidate. In the transformed image section the object’s image has a standard size (measured in „pixels”), independent of distance.

• The transformed image may be used for feature extraction and object recognition; in order to take the distance-dependent variation of image contrast into account certain parameters of fea-ture extraction, such as thresholds, may be ad-justed according to the estimated distance.

Figure 2: Ponzo’s illustra-tion; the two horizontal lines have, in fact, the same length.

• The image transformation is performed by sub-sampling the image in a distance-dependent way. Subsampling is a very fast operation, and it has the additional advantage of reducing the amount of data to be processed in the subse-quent stages of object recognition.

This approach will be described in more detail in the sequel.

4. Distance-Dependent Subsampling

4.1. Concept

Due to our goal is to realize service mobile robots working in office environments. Therefore, this do-main may serve here as an example for introducing the concept.

For the facilitation it is supposed that a camera is mounted in a mobile robot navigating in an office environment. The camera’s optical axis is parallel to the corridor’s floor plane. The robot’s task is, for instance, to recognize objects appearing in the front of the camera without knowledge of the vision sys-tem.

The first subtask is to detect the corridor’s floor edges in the image. The image region corresponding to the corridor’s floor is then searched for objects (It is supposed these tasks were performed by the mo d-ule of the object detector that is not described here).

In reality, the corridor’s floor normally has the same width. Due to perspective distortion, distant sections of the corridor’s floor appear, however, narrower in the image than nearby sections (Fig. 3). In order to compensate this distortion of the apparent corridor’s floor width the image is subsampled in the horizontal row within the image of the corridor’s floor into n equidistant subsegments.

The distance, d, between the subsampled points (Fig. 3) in each line is given by:

where W is the width of the corridor’s floor in that specific image row (Fig. 3). The maximum distance up to which this transformation may be performed is reached when the number subsampled pixels is equal to the number of pixels on the corridor’s floor image in that row.

The horizontal density of subsampled points, and thus n, should be sufficiently high to guarantee that even the narrowest objects that is to be recognized is covered by a sufficient number of such points.

The vertical density of subsampled points, too, de-pends on the goal. It may be determined underlying the assumption that the perspective distortion in a narrow area is approximately zero. That means the length of the corridor’s floor edge segment between a sampling row and the next one should be equal to the distance between the horizontal subsampled points, i.e., d, (Fig. 4).

4.2. Implementation and Results

The described concept was implemented on a real-time vision system and initially tested in real-world office scenes.

Once an object candidate has been found by search-ing the subsampled image of the corridor's floor, a different kind of subsampling is used for analyzing the object candidate. In practice, it is noted that not all of the various parts of the object to be recognized have the same distance from the camera, but these differences are small in comparison with the distance between the object and the camera. Therefore, it is a permissible simplification to consider all visible parts of the object to be located within a common object plane that is parallel to the image plane (Fig. 4). Thus, the subsampling points covering the expected object are arranged in a square grid with a horizontal and vertical period equal to the horizontal subsam-pling period that is applicable to the floor plane where the object plane intersects with the floor plane (Fig. 4). ( )1

1−=

nWd

Figure 3: Concept for transforming the image of the corridor’s floor into an image of distance-invariant width by horizontal sub-sampling (no vertical sampling in this example).

Figure 4: The subsampling interval on the image of the floor plane varies with distance. The subsampling grid in the image region of the expected object is a square con-tinuation of the grid used on the image of the floor plane.

In this way it is assured that the sampling interval in the object plane has a fixed value (a certain fraction of the corridor's floor width), independent of the dis-tance between the object plane and the camera. Con-sequently, the image of a given object always com-prises the same number of subsampled points, re-gardless of the object’s distance from the camera.

Fig. 5 and 6 show examples resulted from real-world experiments. The same object appearing in an office environment was recognized in different positions and distances. It is obvious that the goal of imaging objects in widely varying distances with the same scale factor, and thus in a distance-invariant size, has been accomplished.

Fig. 5: The object was recognized in the near distances, but in different positions.

Fig. 6: The same object from the upper images was recog-nized in far distance and in different positions.

5. Distance Estimation

A prerequisite for distance-dependent subsampling is an estimate of the distance between the camera and objects in the scene. If, as in the previous ex-ample, the objective is object recognition in office environments, the width of the corridor’s floor in the image may used as a clue for estimating depth. The width of a corridor is usually constant over long distances; the width of its image is then in-versely proportional to depth if the viewing direc-tion of the camera is parallel to the floor. In a dif-ferent way of looking at it, the width of the corri-dor may be used as a reference for constructing a local sampling grid that has a fixed period on the (vertical) surface of a phys ical object, regardless of depth.

A similar approach may be taken in some other environments designed for the operation of mobile systems, such as roads or marked lanes on high-ways, canals, or railways.

If a calibrated camera is available the distance from an unknown object resting on the corridor‘s

floor may be calculated from the vertical image coordinate of its lower edge. If the camera is only partially calibrated the vertical image coordinate of an object’s lower edge still gives information on relative depth; in combination with other clues this may allow a distance estimation. If the true size of an object and the imaging scale factor of the cam-era are known depth may be computed from the size of the object’s image. Motion stereo or paral-lax stereo may also be used for depth measurement, and if the robot carries some distance measuring equipment its data may be used, provided that the correspondence between distance data and objects in the camera image can be established.

6. Application

The method of distance-dependent subsampling is applied to realize calibration-free vision-guided mobile robots working in office environments. It was, therefore, tested in a number of applications, including office scenes and object recognition in indoor-scenes.

Figure 7 shows examples of two different mobile robots B21r and Pioneer (the left and right one re-spectively in Figure 7) with quite different physical characteristics and configurations used for our ex-periments.

Fig. 7: The stereo visual B21r (left) and the monocular visual Pioneer robots (right) used for experiments.

Without custom or optional attachments, the Pioneer measures only 18 inches (45 cm) from stem to stem, 14.3 inches (36 cm) across the wheels, and nine inches (22.5 cm) to the top of the console (Fig. 7). While the robot B21r has, e.g., diameter = 52.5 cm, height = 106 cm and weight = 122.5 kg.

Because the corridor’s floor edge recognition is diff i-cult in our office environment, the subsampling is not, based on the floor edges, but on the edges of the skirting board with the wall. That means the subsam-

plingis actually performed on the plane about 10 cm above the floor plane (Fig. 5 and Fig. 6). The object recognition is, therefore, transformed and performed in this plane, too.

Boxes with distinct dimensions have been used for learning and recognizing. The object was located in different distances (nearby, middle and far) in the corridor for learning, and then in other distances for recognizing. Distance-dependent subsampling was used here for creating a database of distance-independent representations of various objects oc-curring in office environments. In a subsequent recognition phase the objects appeared in varying distances in front of a moving robot and were rec-ognized in real time. Recognition was accomp-lished by converting the actual image of the object to a standardized size and comparing a description of this image with the image descriptions in the data base.

7. Summary

A novel approach for compensating perspective dis-tortion by a depth-dependent image transformation has been introduced. Its effectiveness in recognizing objects that may appear in unpredictable and widely varying distances has been demonstrated by imple-menting it on a real-time vision system and testing it in real-world experiments.

References

[1] J. J. Gibson, "The Perception Of The Vi sual World.'' Houghton Mifflin, Boston, MA, 1950.

[2] V. Graefe, "Calibration-Free Robots.” The 9th Intelligent System Symposium. Japan Society of Mechanical Engineers, pp. 27-35, Japan, 1999.

[3] I. Horswill, “Visual Collision Avoidance,” Proc. Of IEEE/RSJ/GI International Conference on Intelligent Robots and Systems, IROS’94, Mu-nich, pp. 902-909, 1994.

[4] M.-C. Nguyen, “Object Manipulation by Cali-bration-Free Vision-Guided Robots,” Disserta-tion, Bundeswehr University Munich, 2000.

[5] M.-C. Nguyen, V. Graefe, ``Visual Recognition of Objects for Manipulating by Calibration-Free Robots,'' In Kenneth, V. T., et al. (eds.): Machine Vision Applications in Industrial Inspection VIII. Proc. of IST/SPIE, Vol. 3966, San Jose, ISBN 1-902856-02-3, pp. 290-298, 1999.

[6] U. Regensburger, “Zur Erkennung von Hin-dernissen in der Bahn eines autonomen Stras-senfahzeugs durch maschinelles Echzeitsehen.“, Dissertation, Bundeswehr University Munich, 1994.

[7] U. Regensburger, V. Graefe, „Visual Recogni-tion of Obstacles on Roads.“ In V. Graefe (ed.): Intelligent Robots and Systems, Amsterdam: El-sevier, pp. 73-86, 1994

[8] I. Rock, “Wahnemung: vom vis uellen Reiz zum Sehen und Erkennen.”, Spektrum der Wissen-schaft Verlagsgesselschaft , Heidenberg, 1985.

0,1(�&/($5$1&(�52%27�+$1'�)25�$�'(0,1,1*�52%27� ���

7\WXV� :2-7$5$ �� .HQ]R� 121$0, ��

� W\WXV#PHF��WP�FKLED�X�DF�MS�&KLED� 8QLY��� *UDGXDWH� 6FKRRO� RI� 6FLHQFH� DQG� 7HFKQRORJ\�� � QRQDPL#PHFWK�WP�FKLED�X�DF�MS� &KLED� 8QLY��� 'HSW�� RI� (OHFWURQLFV� DQG� 0HFKDQLFDO� (QJLQHHULQJ��

���

$EVWUDFW� ��7KLV� SDSHU� SURSRVHV� D� PLQH� FOHDUDQFH� KDQG� WR� XVH�ZLWK�WKH�PLQH�FOHDUDQFH�URERW�&20(7�GHYHORSHG�DW�WKH� &KLED� 8QLYHUVLW\�� 7KH� URERW� KDQGV� WDVN� LV� WR�QHXWUDOL]H�GLVDUP� DQG� UHPRYH� DUPHG� PLQHV� RQ� DQ�DFWXDO� PLQH� ILHOG�� 7KH� URERW� KDQGV� GHVFULSWLRQ� DQG�DSSOLFDWLRQ� LV� GLVFXVVHG�� $OVR� D� IRUFH�VHQVRU�OHVV�KDQG�DSSURDFK�LV�GHVFULEHV�EULHIO\��

.H\ZRUGV� �� URERW�� KDQG�� PLQH� FOHDUDQFH�� WHOH�RSHUDWLRQ��

���,QWURGXFWLRQ��7KHUH�DUH�PDQ\�PLQHILHOGV�DOO�RYHU� WKH�ZRUOG��0RUH�WKDQ� RQH� KXQGUHG� PLOOLRQ� DQWL�SHUVRQQHO� PLQHV� VWLOO�OD\� XQGHUJURXQG�� (YHU\� \HDU� DERXW� ������ SHRSOH�EHFRPH� YLFWLPV� RI� PLQHV�� 0LQHV� GHOD\� SHDFH�SURFHVVHV�DQG� HFRQRPLF� UHFRQVWUXFWLRQ��7KH\�PDNH�WKH� ODQG� XVHOHVV� IRU� DJULFXOWXUH� DQG� IRU� OLYLQJ��0LQH�VXUYLYDOV� DUH� QRW� DEOH� RU� QRW� IXOO\� DEOH� WR� ZRUN� LQ�RUGHU� WR� VHFXUH� WKH� H[LVWHQFH� RI� WKHLU� IDPLOLHV� DQG�RIWHQ�WKH\�HYHQ�QHHG�WKH�VXSSRUW�RI�WKHLU�IDPLOLHV�� �

1RZDGD\V� WKH� PLQH� FOHDUDQFH� SURFHVV� LV� PRVWO\�FDUULHG�RXW�PDQXDOO\��)LJ����ZKLFK� LV�YHU\�VORZ�DQG�GDQJHURXV�� 8VLQJ� WKH� SUHVHQWO\� DSSOLHG� PHWKRGV� WR�FOHDU� DOO� WKH� PLQHV� LW� ZLOO� WDNH� EHWZHHQ� VRPH�KXQGUHGV�DQG�RQH�WKRXVDQG�\HDUV�� �

8VLQJ�D�URERW�LQVWHDG� WKH�PLQH�FOHDUDQFH�DFWLRQV�FDQ�EH� VSHHG� XS� DQG� PDNH� GHPLQHU¶V� ZRUN� VDIHU�� $�EURNHQ�SDUW�RI�D�URERW�FDQ�EH�UHSODFHG�ZKLOH�LQMXUHG�SDUWV�RI�WKH�ERG\�RI�D�GHPLQLQJ�SHUVRQ�FDQQRW��

:LWK�WKLV�PRWLYDWLRQ�WKH�PLQH�FOHDUDQFH�KDQG�SURMHFW�LV�FDUULHG�RXW�� �

7KH� PLQH� FOHDUDQFH� KDQG� ZLOO� EH� LQVWDOOHG� RQ� WKH�DOUHDG\�H[LVWLQJ�URERW�&20(7�,,,��)LJ�����7KH�PLQH�GHWHFWLRQ�URERW�&20(7�,,,� LV�D���OHJ�ZDONLQJ�URERW�ZLWK� �� DUPV� LQ� WKH� IURQW�� ,W� LV� GULYHQ� E\� K\GUDXOLF�SRZHU�� 7KH� URERW� ZHLJKV� ���NJ�� LWV� OHQJWK� LV� �P��ZLGWK� ���P� DQG� WKH� KHLJKW� ���P�� ,Q� DUHDV� ZKHUH�PLQHV� DUH� H[SHFWHG� WKH� URERW� ZDONV� ZKLOH� LQ� VDYH�DUHDV�WKH�URERW�PRYHV�XVLQJ�D�FUDZOHU�ZKLFK�LV�IDVWHU�WKDQ�PRYLQJ�E\�OHJV�� � �

$� PLQH� GHWHFWLRQ� VHQVRU� LQVWDOOHG� RQ� RQH� DUP� LV�GHWHFWLQJ� WKH� PLQH� SRVLWLRQ� DQG� WKH� RQERDUG�FRPSXWHU� LV� DEOH� WR� GUDZ� D� PDS� RI� WKH� PLQHILHOG�XVLQJ�D�*36�� �

)LJ���0DQXDO�GHPLQLQJ��GHPLQHU�DW�ZRUN�

$V� DQ� DGGLWLRQ� WR� WKH� FXUUHQW� FRQILJXUDWLRQ� D� PLQH�FOHDUDQFH�KDQG�ZLOO�EH�LQVWDOOHG�RQ�WKH�URERW��

7KH� KDQG�ZLOO� EH� XVHG� WR� UHPRYH� EXULHG�PLQHV�� FXW�WULS�ZLUHV�DV�ZHOO�DV�WR�GLVDUP�PLQHV�� +ROGLQJ� WRROV�WKH� KDQG� ZLOO� EH� DEOH� WR� UHPRYH� REVWDFOHV�� FXW�YHJHWDWLRQ�RU�GLJ�D�KROH�LQ�WKH�JURXQG��

�

)LJ���$FWXDO�URERW�&20(7�,,,�

7KH�DUWLFOH�LV�VWUXFWXUHG�DV�IROORZV��)LUVW��RWKHU�URERW�KDQG�SURMHFWV�DUH�VXUYH\HG�EULHIO\��1H[W��GHVFULSWLRQ�

RI� PLQHILHOG� HQYLURQPHQW� DQG� PLQH� W\SHV� LV� JLYHQ��1H[W�� WKH�PLQLPDO� UHTXLUHPHQWV� IRU� PLQH� FOHDUDQFH�DUH� GLVFXVVHG�� 7KH� JHQHUDO� FRQFHSW� RI� WKH� XVDJH� RI�WKH� KDQG� ZLWK� FRRSHUDWLRQ� ZLWK� WKH� 5RERW� DQG� D�UHPRWH�FRQWURO�VWDWLRQ�LV�GHVFULEHG��7KH�VWUXFWXUH�DQG�GHVFULSWLRQ� RI� WKH� KDQG� DV� ZHOO� DV� WKH� XVDJH� RI�VHQVRUV� LV� SUHVHQWHG� LQ� WKH� QH[W� SDUW� DQG� DW� ODVW� WKH�FRQWURO�VWUDWHJ\�LV�GHVFULEHG��2WKHU�+DQG�5HVHDUFKHV�0DQ\� UHVHDUFKHUV� FRQFHUQHG� ZLWK� URERW� KDQGV� DUH�VWXG\LQJ� JUDVSLQJ� DQG� PDQLSXODWLRQ� RI� GLIIHUHQW�XVXDOO\� OLPLWHG� NLQGV� RI� REMHFWV� LQ� D� VSHFLDOO\�SUHSDUHG� HQYLURQPHQW�� 0DQ\� UHVHDUFKHUV� DUH�FRQFHQWUDWHG� RQ� WKH� LPLWDWLQJ� RI� WKH� KXPDQ� KDQG��0DQ\� DQWKURSRPRUSKLF� KDQGV� ZHUH� GHYHORSHG��6WDQIRUG�-3/�KDQG�� 8WDK�0,7�KDQG�� %HOJUDGH�86&�KDQG�� 2PL�KDQG� �1$6$��� '5/�,,�+DQG� �*HUPDQ�$HURVSDFH� &HQWHU��� ',67�+DQG� �8QLY�� RI� *HQRYD��DQG� 3QHXPDWLF� $QWKURSRPRUSKLF� KDQG� >�@�� $OVR�PDQ\� QRQ�DQWKURSRPRUSKLF� KDQGV� ZHUH� GHYHORSHG��.DUOVUXKH�KDQG� >�@�� <DVXNDZD�KDQG�� 7RN\R�8QLY��KDQG�� 8QLY�� RI� 0DOWD� KDQG� >�@� DQG� RWKHUV�� 7KH�UHIHUHQFH� >�@�JLYHV�D�JRRG�RYHUYLHZ�RI� WKH�GLIIHUHQW�URERW� KDQG� W\SHV� DQG� ILHOGV� RI� UHVHDUFK� FRQFHUQLQJ�URERW� KDQG� PDQLSXODWLRQ�� 5HVHDUFK� RQ� GH[WHURXV�KDQGV� IRU� D� VSHFLDO� HQYLURQPHQW� LV� FDUULHG� RXW� IRU�H[DPSOH� LQ� WKH� SURMHFW��$0$'(86�'H[WHURXV� 6XE�6HD�+DQG� >�@��7KHUH�KDV�DOVR�EHHQ�GHVLJQHG�D� URERW�KDQG�ZKLFK� LV�SODQQHG� WR�EH�XVHG�ZLWK�D�KRXVHZRUN�URERW�>�@��8QWLO�QRZ�WR�RXU�EHVW�NQRZOHGJH�WKHUH�KDV�QRW� EHHQ� SURSRVHG� D� URERW� KDQG� IRU� WKH� H[FOXVLYH�XVDJH�DV�PLQH�FOHDUDQFH�URERW�KDQG��&RQFHUQLQJ� WKH� VWUXFWXUH� WKHUH�DUH�GLIIHUHQW�NLQGV�RI�DSSURDFKHV��7HQGRQ�GULYHQ�KDQGV�KDYH�WKH�DGYDQWDJH�RI�PRYLQJ�WKH�XVXDOO\�KHDY\�PRWRUV�WR�XSSHU�SDUWV�RI�WKH�DUP�UHGXFLQJ�WKH�LQHUWLD�RI�WKH�PDQLSXODWRU�KDQG�V\VWHP�� 1RZDGD\V� KRZHYHU� PDQ\� TXLWH� VPDOO� DQG�OLJKW�EXW�VWLOO�SRZHUIXO�PRWRUV�DUH�DYDLODEOH�DQG�DV�IDU�DV�WKH�KXPDQ�KDQG�LV�QRW�WKH�DLP�FRQVWUXFWLQJ�D�KDQG�ZLWK�PRWRUV�ORFDWHG�GLUHFWO\�LQVLGH�WKH�KDQG�VWUXFWXUH�LV�SRVVLEOH��0DQ\�UHVHDUFKHUV�XVH�HOHFWULF�PRWRUV��EXW�DOVR� K\GUDXOLF�� SQHXPDWLF�� DQG� VRPH� RULJLQDO�DFWXDWRUV�OLNH�DFWXDWRU�EDVHG�RQ�PHWDOV�ZLWK�PHPRU\��'/5�KDQG���XOWUDVRQLF�PRWRUV� �7R\DPD�/DE���� IOXLG�ILOOHG� WHQWDFOHV� >�@�� DLU�PXVFOHV� �7KH�6KDGRZ�5RERW�&RPS���DUH�EHLQJ�H[SORUHG������0LQH� )LHOG� (QYLURQPHQW�������*HQHUDO�GHVFULSWLRQ�RI�PLQH�ILHOG�HQYLURQPHQW�7KHUH�DUH�D� ORW� RI�PLQH�ILHOGV�LQ�PDQ\�FRXQWULHV��7KH�HQYLURQPHQW� LV� YHU\� GLIIHUHQW� GHSHQGLQJ� RQ� WKH�JHRJUDSKLF� ORFDWLRQ�� :H� KDYH� WR� GHDO� ZLWK�HQYLURQPHQWV� OLNH� GHVHUW�� MXQJOH�� IRUHVW�� EXVK� ODQG��KLJK�JUDVV��HYHQ�DQG�XQHYHQ��VRIW�RU�KDUG��ZHW�RU�GU\�WHUUDLQ�� KLJK� RU� ORZ� WHPSHUDWXUH�� 7KH� URERW� DQG� LWV�PDQLSXODWRU�KDYH� WR�EH�DEOH� WR�ZRUN�DOVR�GXULQJ� WKH�UDLQ�DQG�LQ�WKH�QLJKW��7KH�URERW�IDFHV�PDQ\�REVWDFOHV�

OLNH�VWRQHV��EURNHQ�WUHHV�DQG�WUHH�EUDQFKHV��7KH�URERW�KDQG� KDV� WR� EH� DEOH� WR� SURYLGH� REVWDFOH� UHPRYDO��FXWWLQJ�RI�JUDVV��EXVKHV�DQG�VPDOO�WUHHV��7KH�GLIIHUHQW�WRROV� WKDW� WKH� KDQG� FDQ� XVH� KDYH� WR� EH� DYDLODEOH�TXLFNO\��,QVWHDG�RI�PDQ\�DUPV�RU�PDQ\�HQG�HIIHFWRUV�LQVWDOOHG�RQ�RQH�DUP�D�URERW�KDQG�JUDVSLQJ�GLIIHUHQW�WRROV� ZLWK� D� FRUUHVSRQGLQJ� JULS� LV� SODQQHG��7KH�SUHVHQW� UHVHDUFK� LV� FRQFHQWUDWHG� RQ� WKH� PLQH�FOHDUDQFH�LQ�$IJKDQLVWDQ�ZKHUH�WKH�VRLO�LV�PRVWO\�IODW��YHU\�KDUG�DQG�GU\��7KH�WHPSHUDWXUHV�DUH�KLJK�DQG�WKH�RSHUDWLRQ�LQ�ERWK�GU\�DQG�UDLQ\�VHDVRQ�LV�UHTXLUHG��

)LJ���'LIIHUHQW�PLQH�W\SHV�

�

0LQH�7\SHV�,Q� WKH� PLQHILHOG� RQH� FDQ� ILQG� PDQ\� GLIIHUHQW� PLQH�W\SHV� �)LJ���� DQG� 8;2V� �XQH[SORGHG� RUGLQDQFHV���*HQHUDOO\� WKHUH� DUH� DQWLSHUVRQQHO� PLQHV� �$3�� DQG�DQWLWDQN� PLQHV� �$7��� $7V� DUH� XVXDOO\� ELJJHU� DQG�KHDYLHU�� 7KH\� DUH� EXULHG� DW� D� ELJJHU� GHSWK� DQG� WKH�WKUHVKROG� ZHLJKW� WR� WULJJHU� WKHP� LV� KLJKHU� WKDQ� $3�DQG� OLHV� XVXDOO\� LQ� WKH� UDQJH� �������� NJ�� $7V� DUH�XVXDOO\�HDV\� WR�GHWHFW�ZLWK�D�PHWDO�GHWHFWRU�EHFDXVH�WKH\� FRQWDLQ� D� ORW� RI� PHWDO�� $3¶V� DUH� VPDOOHU� DQG�OLJKWHU��7KH\�DUH�VRPHWLPHV�EXULHG�VRPH�FHQWLPHWHUV�EHORZ� WKH� JURXQG� VXUIDFH� DQG� VRPHWLPHV� WKH\� HYHQ�OLH�RQ� WKH�JURXQG�VXUIDFH��7KH�GHWHFWLRQ�RI�D�SODVWLF�$3� PLQH� LV� GLIILFXOW�� :H� FODVVLI\� $3� LQ� EODVW��IUDJPHQWDWLRQ�� GLUHFWLRQDO� DQG�ERXQFLQJ� W\SHV��%ODVW�PLQH�W\SHV�DUH�RIWHQ�PDGH�RI�SODVWLF�DQG�DUH�WKHUHIRUH�GLIILFXOW� WR�GHWHFW��)UDJPHQWDWLRQ�PLQHV� FRQVLVW� RI� D�PHWDO�KRXVLQJ�WKDW�LV�EUHDNLQJ�LQWR�SLHFHV�GXULQJ�WKH�H[SORVLRQ��7KHLU�PHWDO�FRQWHQW�PDNHV� WKHP�HDVLHU� WR�GHWHFW��7KH\�DUH�XVXDOO\�RQ�WKH�JURXQG�VXUIDFH�VR�WKDW�HYHQ� DQ� RSWLFDO� GHWHFWLRQ� LV� VRPHWLPHV� SRVVLEOH��'LUHFWLRQDO�PLQHV�DUH�PRXQWHG�DERYH� WKH�JURXQG�RQ�VWLFNV�WKDW�PDNHV�WKHP�HDVLHU�WR�GHWHFW�RSWLFDOO\�� �

0RVW�PLQHV�DUH�WULJJHUHG�ZKHQ�SXVKHG�IURP�WKH�WRS��7KHUH�DUH�DOVR�PLQHV�XVLQJ�WULS�ZLUH�WKDW�LV�D�GLIIHUHQW�NLQG�RI� WULJJHULQJ� D�PLQH�ZKLFK� UHTXLUHV� D�GLIIHUHQW�GLVDUPLQJ� KDQGOLQJ�� 7KHUH� DUH� DOVR� RWKHU� W\SHV� RI�WULJJHULQJ�� 7KH� SUHVHQW� UHVHDUFK� LV� FRQFHQWUDWLQJ� RQ�

KDQGOLQJ� WKH� PRVW� FRPPRQ� PLQH� W\SHV� IRXQG� LQ�$IJKDQLVWDQ�ZKLFK�DUH�WKH�$3�320=��� �LQ� WKH�)LJ���WKH� ORQJ� RQH�� 301� �LQ� WKH� SLFWXUH� EHORZ� WKH� ULJKW�320=���PLQH��� 301��� �QRW� LQ� WKH� SLFWXUH�� WKH�$7�70�VHULHV��LQ�WKH�SLFWXUH�LQ�WKH�XSSHU�ULJKW�FRUQHU���

�����6SHFLDO�,VVXHV�RI�0LQH�&OHDUDQFH�0LQLPDO�5HTXLUHPHQWV�7R� PDNH� PLQH� FOHDUDQFH� D� VDWLVI\LQJ� WDVN� IRU� DQ�DFWXDO� PLQH� ILHOG� WKH� IROORZLQJ� EDVLF� UHTXLUHPHQWV�KDV�WR�EH�VDWLVILHG�� �

����&OHDUDQFH�RI�WKLFN�YHJHWDWLRQ��OLNH�LQ�D�MXQJOH��

����ORZ�JUDVV�FXWWLQJ�

����FXWWLQJ�WULS�ZLUH�

����$7�UHPRYDO����KDQGV��

����8;2�UHPRYDO����KDQGV��

7KLV� SURMHFW� LV� DLPLQJ� DW� VDWLVI\LQJ� DOO� RI� WKHVH�UHTXLUHPHQWV�VWHS�E\�VWHS�� �

*UDVS�7\SHV�$�KXPDQ�LV�XVLQJ�PDQ\�JUDVS�W\SHV�GHSHQGLQJ�RQ�WKH�REMHFW�VL]H��VKDSH�DQG�PDWHULDO��6RPH�RI�WKHP�ZLOO�EH�LPSOHPHQWHG�RQ� WKH�PLQH�FOHDUDQFH� URERW�KDQG��7KH�QHHGHG� JUDVS� W\SHV� DUH� GHSHQGHQW� RQ� WKH� WDVN� WR� EH�FDUULHG�RXW��7KH�UHTXLUHPHQWV� IRU� WKH�URERW�KDQG�DUH�DV�IROORZV��

�� KROGLQJ�DQ�$3��$7��8;2�

�� KROGLQJ�D�WULSZLUH�D�WULS�ZLUH�LV�RIWHQ�D�WKLQ�DOPRVW�LQYLVLEOH�ZLUH�

�� FXWWLQJ�WULS�ZLUH�

�� FKDQJLQJ�WRROV�DQG�KROGLQJ�WKHP�

�� FXW�WKH�JUDVV�

�� FXWWLQJ�VPDOO�WUHHV��EDPERRV�DQG�EXVKHV�

�� GLJJLQJ�WKH�VRLO�ZLWK�D�VKRYHO�

�� EORZLQJ� WKH� VRLO� DZD\�ZLWK�KLJK�VSHHG�SOHDVXUH�DLU�DQG�KLJK�SUHVVXUH�ZDWHU�

�� KROGLQJ�D�SURGGHU�WR�H[DPLQH�VRLO¶V�VWLIIQHVV�

�� PDUNLQJ�PLQH�SRVLWLRQ�

7R� DFKLHYH� WKLV� DLP� ZH� KDYH� WR� LPSOHPHQW� D�SUHFLVLRQ�JUDVS�DQG�D�SRZHU�JUDVS�>�@��7KH�SUHFLVLRQ�JUDVS� D� JUDVS� W\SH� ZKHUH� RQO\� WKH� ILQJHUWLSV� DUH�KROGLQJ�WKH�REMHFW��7KLV�NLQG�RI�JUDVS�SXWV�HPSKDVLV�RQ� GH[WHULW\� DQG� VHQVLWLYLW\�� 7KH� SRZHU� JUDVS� LV� D�

NLQG�RI�JUDVS�ZKLFK�DOVR� LQYROYHV� WKH� LQQHU�DUHDV�RI�WKH� ILQJHU� GLJLWV� WR� KROG� WKH� REMHFW��3RZHU� JUDVS� LV�PRUH� VHFXUH�DQG�VWDEOH� WKDQ� WKH�SUHFLVLRQ�JUDVS��7R�KROG�WKH�WULSZLUH�IRU�H[DPSOH�WKH�SUHFLVLRQ�SULVPDWLF�JUDVS�ZLOO�EH�LPSOHPHQWHG��7R�KROG�PLQHV��REVWDFOHV�DV� ZHOO� DV� WRROV� D� SULVPDWLF� SRZHU� JUDVS� DQG� D�FLUFXODU�SRZHU�JUDVS�ZLOO�EH�LPSOHPHQWHG�� �

�����,PDJH�RI�WKH�+DQG�DQG�&20(7�,,,�6WUXFWXUH�7KH�KDQGV� VL]H� LV�SODQQHG� WR�EH� WZLFH�RI� WKH�KXPDQ�KDQG� DQG� LWV� ZHLJKW� DERXW� �� NJ�� 7KH� KDQG� ZLOO�EH�DWWDFKHG� WR� WKH� URERW� ERG\� E\� D� ��'�2�)�� DUP�� 7KH�KDQG� �)LJ���� LWVHOI� ZLOO� KDYH� �� ILQJHUV� ZLWK� HDFK� ��'�2�)��7KH�KDQG�KDV�DOVR�DQ� LQWHJUDWHG�ZLUH�FXWWLQJ�GHYLFH�� D� KLJK�VSHHG� SUHVVXUH� DLU� DQG� D� KLJK� VSHHG�ZDWHU� SLSH�� 7KH�ZLUH� FXWWHU�ZLOO� EH� XVHG� WR� FXW� WULS�ZLUH��,Q�FRQQHFWLRQ�ZLWK�FRORU�SDLQW�WKH�KLJK�SUHVVXUH�DLU�ZLOO�EH�XVHG�WR�PDUN�GDQJHURXV�VSRWV�GHWHFWHG�E\�WKH� PLQH� GHWHFWLRQ� VHQVRU�� 7KH�KLJK�SUHVVXUH�DLU�>�@�ZLOO�EH�XVHG�VRPHWLPHV�WR�H[FDYDWH�WKH�PLQH�DQG�DOVR�WR�IUHH�WKH�PLQH�VXUIDFH�IURP�WKH�VRLO�� )RU� KDUG� VRLO�WKH�KLJK�VSHHG�ZDWHU�ZLOO�EH�XVHG�WR�FXW�WKH�LW��,I�WKH�PLQH�VXUIDFH�LV�IUHH�WKH�PLQH�W\SH�FDQ�EH�UHFRJQL]HG�E\� DQ� RSHUDWRU� WKURXJK� D� FDPHUD� LQVWDOOHG� RQ� WKH�URERW� DQG� �ODWHU� E\� WKH� DXWRPDWLF� RSWLFDO� PLQH�GHWHFWLRQ� V\VWHP���7KH� NQRZOHGJH� RI� WKH�PLQH� W\SH�VWRUHG�LQ�D�GDWDEDVH�LV�HTXLYDOHQW�ZLWK�WKH�NQRZOHGJH�RI� WKH� H[DFW� VL]H�� WKH� ZHLJKW�� WKH� FRQWHQW� DQG� WKH�PDWHULDO�RI�WKH�KRXVLQJ��

2XU� URERW� KDQG� LV� GULYHQ� E\� HOHFWULF� PRWRUV��*HQHUDOO\� URERW� KDQGV� IRU� ILQJHU� WLS� JUDVS� DUH�HTXLSSHG� ZLWK� IRUFH� VHQVRUV� RQ� WKH� ILQJHU� WLSV� DQG�KDQGV� IRU� SRZHU� JUDVSV� DUH� HTXLSSHG� ZLWK� WRXFK�VHQVRUV� RQ� WKH� LQQHU� VLGH� RI� WKH� ILQJHU� VLGH� DQG� WKH�SDOP��6LQFH�WKHUH�DUH�GLIIHUHQW�W\SHV�RI�REMHFWV�ZLWK�D�SULRUL� XQNQRZQ� VKDSH� D� KDQG� ZLWK� VHQVRUV� FDQ�KDSSHQ� WR� EH� EORFNHG� E\� WKH� REMHFW� WR� EH� JUDVSHG�ZLWKRXW� WKDW� WKH� VHQVRUV� ZRXOG� UHJLVWHU� LW�� ,Q� RXU�DSSURDFK� ZH� GRQ¶W� XVH� IRUFH� VHQVRUV� DQG� WRXFK�VHQVRUV��:H�DUH�FDUU\LQJ�RXW�EDVLF�UHVHDUFK�WKDW�DLPV�DW�GHYHORSLQJ�D�KDQG�XVLQJ�RQO\�MRLQW�PRWRUV�SRVLWLRQ�VHQVRUV�WR�HVWLPDWH�WKH�IRUFH�DFWLQJ�RQ�WKH�ILQJHUV��

�

)LJ���0LQH�FOHDUDQFH�URERW�KDQG�

�

0DQLSXODWRU�7KH�PDQLSXODWRU�WR�ZKLFK�WKH�KDQG�LV�DWWDFKHG�KDV���'�2�)�� WR� JLYH� WKH� KDQG� DQ\� DUELWUDU\� SRVLWLRQ� DQG�RULHQWDWLRQ�� 7KH� PDQLSXODWRU� H[FOXGLQJ� WKH� KDQG�ZULVW�LV�GULYHQ�E\�K\GUDXOLF�SRZHU��7KH�ZULVW�LV�XVLQJ�HOHFWULF�PRWRUV��

���*HQHUDO� FRQFHSW� RI� 0LQH� &OHDUDQFH� +DQG��7KH�URERW�PLQH�FOHDUDQFH�KDQG�FRQWURO�LV�SODQHG�WR�EH�SDUWO\� DXWRQRPRXV� DQG� UHPRWH� FRQWUROOHG��:LWK� WKH�KHOS� RI� WKH� FDPHUD� LQVWDOOHG� RQ� WKH� URERW� WKH� YLVXDO�GHWHFWLRQ� RI� REVWDFOHV�� VXUIDFH� ODLG� PLQHV� RU� PLQHV�XVLQJ� WULS� ZLUH� LV� SRVVLEOH�� 7KH� FRQFHSW� RI� PLQH�FOHDUDQFH�LV�SODQQHG�WR�EH�DV�IROORZV��7ZR�URERWV�RI�WKH� W\SH� &20(7� DQG� D� PLQH� FOHDUDQFH� YHKLFOH� DUH�ZRUNLQJ� WRJHWKHU��7KH� ILUVW� URERW� LV�HTXLSSHG�ZLWK�D�PLQH� GHWHFWLRQ� VHQVRU� DQG� D� PLQH� FOHDUDQFH� KDQG��)LJ�����7KH�VHFRQG�URERW�LV�HTXLSSHG�ZLWK�WZR�PLQH�FOHDUDQFH�KDQGV� �)LJ�����7KH�PLQH� FOHDUDQFH�YHKLFOH�LV�XVHG�RQO\�WR�FOHDU�$3�PLQHV��7KH�ILUVW�URERW¶V�WDVN�LV� WR� GHWHFW� PLQHV�� PDUN� WKHLU� SRVLWLRQ� DQG� FODVVLI\�WKHP� LQWR� WZR� JURXSV�� $3� DQG� $7�8;2�� 7KH� ILUVW�URERW� SURYLGHV� DOVR� REVWDFOH� FOHDUDQFH�� 7KH� VHFRQG�URERW� LV� D� PLQH� FOHDUDQFH� URERW� VSHFLILHG� WR� FOHDU�PDLQO\�$7V�DQG�8;2V��7KH�PLQH� FOHDUDQFH�YHKLFOH�UHPRYHV� WKH� XSSHU� VRLO� OD\HU� DQG� VHSDUDWHV� PLQHV�IURP� HDUWK�� ,W� FOHDUV� RQO\� $3V� DYRLGLQJ� $7V� DQG�8;2V� GHWHFWHG� E\� WKH� ILUVW� URERW�� 7KH� YHKLFOH� LV�UREXVW�HQRXJK�QRW�WR�EH�GDPDJHG�E\�DQ�H[SORGLQJ�$3��,WV� DGYDQWDJH� LV� WKH� KLJK� PLQH� FOHDUDQFH� VSHHG�� ,Q�SODFHV�WKDW�WKH�PLQH�FOHDUDQFH�YHKLFOH�FDQ¶W�UHDFK�RU�LQ�FDVHV� WKH�PLQH�FODVVLILFDWLRQ� WR�$3�DQG�$7�8;2�JURXS� LV� QRW� SRVVLEOH�� WKH� URERWV� SURYLGH� WKH� $3�FOHDUDQFH��

:H�GLVWLQJXLVK� EHWZHHQ� REVWDFOH� JUDVSLQJ� DQG�PLQH�JUDVSLQJ��$Q�REVWDFOH� LV�XVXDOO\�DQ�XQNQRZQ�REMHFW��,WV� VL]H� DQG� ZHLJKW� FDQ� EH� RQO\� URXJKO\� HVWLPDWHG�IURP�WKH�LPDJH�UHFHLYHG�IURP�WKH�FDPHUD��$�PLQH�LV�D�NQRZQ�REMHFW��$V�VRRQ�DV�D�PLQH�FDQ�EH�FODVVLILHG�RSWLFDOO\� E\� W\SH� LW� EHFRPHV� D� NQRZQ� REMHFW�� ,WV�VKDSH��VL]H��ZHLJKW��KRXVLQJ�PDWHULDO�DQG�FRQWHQWV�DUH�NQRZQ�� $� PLQH� W\SH� GDWD� EDVH� FRQWDLQV� DQ�DSSURSULDWH� JUDVS� IRU� HYHU\�PLQH� W\SH��7KH� JUDVS� LV�VSHFLILHG�E\�WKH�SRVLWLRQ�RI�WKH�ILQJHUV�DQG�WKH�IRUFH�QHFHVVDU\�WR�KROG�WKH�PLQH��

7KH�QXPEHU�RI�PLQHV�LV�D�OLPLWHG�VHW��,Q�VRPH�UHJLRQV�RQO\� D� IHZ�PLQH� W\SHV� OD\�XQGHUJURXQG��%HFDXVH�RI�WKLV� IDFW� WKH� UHODWLYHO\� VPDOO�PLQH� W\SH�GDWDEDVH�FDQ�EH�XVHG�IRU�JUDVS�SODQQLQJ��

2EVWDFOHV�$YRLGDQFH�DQG�5HPRYDO��,Q�FDVH�RI�DQ�REVWDFOH�GHSHQGLQJ�RI�VL]H�DQG�NLQG�RI�WKH�REVWDFOH� WKH�URERW�KDQGOLQJ�GLIIHUV�� ,I� WKH�VL]H�RI�WKH�REVWDFOH� LV� VPDOO� �XQHYHQ�JURXQG�� WKH� URERW�ZLOO�GHFLGH� WR� VWHS� RYHU� WKH� REVWDFOH�� ,I� WKH� VL]H� RI� WKH�REVWDFOH�LV�YHU\�ELJ��WUHH��URFN��GHHS�KROH��WKH� URERW�ZLOO�JR�DURXQG�WKH�REVWDFOH��,Q�VRPH�FDVHV�LW�LV�EHWWHU�

WR�UHPRYH� WKH�REVWDFOH�HVSHFLDOO\� LI� WKH\�DUH�ELJ�EXW�QRW� YHU\� KHDY\�� ,I� SRVVLEOH� WKH� REVWDFOH� ZLOO� EH�SXVKHG� DVLGH� �GU\� EXVK� RU� SDUW� RI� D� WUHH�� DQ� HPSW\�EDUUHO��� ,Q� FDVH� WKH� REMHFW� FDQQRW� EH� SXVKHG� DVLGH�VRPH�RWKHU�DFWLRQV�KDV�WR�EH�WDNHQ��7KH�REVWDFOH�FDQ�EH�JUDVSHG�ILUVW�DQG�WKHQ�SXW�DVLGH��ROG�WUHH�EUDQFKHV���8VLQJ� D� FKDLQ� VDZ� VRPH� REVWDFOHV� �VPDOO� WUHHV��EDPERR��FDQ�EH�FXW�DQG� WKHQ�JUDVSHG�DQG�SXW�DVLGH��6RPH�RWKHU�W\SHV�RI�REVWDFOH��OLNH�KLJK�JUDVV��FDQ�EH�FXW�XVLQJ�D�JUDVV�FXWWHU��7KH�JUDVSLQJ�RI�DQ�REVWDFOH�LV� JUDVSLQJ� RI� DQ� XQVWUXFWXUHG� REMHFW�� 7KH� JUDVS�SODQQLQJ� ZLOO� LQFOXGH� SDUDPHWHUV� OLNH� JUDVS� W\SH��F\OLQGULFDO�� FLUFXODU� DQG� VSKHULFDO��� JUDVS� VL]H� DQG�WKH�JUDVS�IRUFH��7KH� WUDMHFWRU\�SODQQLQJ�ZLOO�KDYH� WR�WDNH�LQWR�DFFRXQW�WKH�VL]H�RI�WKH�REMHFW�DQG�LWV�VKDSH��7KH�KROGLQJ�RI�DQ�REVWDFOH�LV�QRW�D�YHU\�FUXFLDO�WDVN��7KH� VOLSSLQJ� RI� WKH� REMHFW� RXW� RI� WKH� JUDVS� DQG� WKH�KLWWLQJ� WKH� JURXQG� E\� WKH� REMHFW� ZLOO� QRW� FDXVH� DQ\�H[SORVLRQ�VLQFH�RQO\�OLJKW�REMHFWV�DUH�JUDVSHG�� �

)LJ���0LQH�GHWHFWLRQ�URERW�&200(7�,,,�HTXLSSHG�ZLWK�D�PLQH�GHWHFWLRQ�VHQVRU�DQG�D�PLQH�FOHDUDQFH�

KDQG�

�

)LJ����0LQH�FOHDUDQFH�URERW�HTXLSSHG�ZLWK�WZR�PLQH�FOHDUDQFH�KDQGV�

0LQH�&OHDUDQFH��,Q�FDVH� WKH�YLVXDO� UHFRJQLWLRQ�GHWHFWV�D� WULS�ZLUH� WKH�ZLUH�ZLOO�EH�WUDFNHG�YLVXDOO\�ERWK�GLUHFWLRQV�XQWLO� LWV�HQGV�� 7KH� PLQH� W\SH� ZLOO� EH� FODVVLILHG� LI� SRVVLEOH��)URP� WKH�PLQH� W\SH� GDWDEDVH� LQIRUPDWLRQ� OLNH� VL]H��ZHLJKW��VKDSH��KRXVLQJ�PDWHULDO��FRQWHQW�RI�WKH�PLQH�DQG�WKH�PRVW�FRPPRQ�IXVH�W\SH�FDQ�EH�H[WUDFWHG��$V�QH[W� DFWLRQ� WKH� ZLUH� ZLOO� EH� DSSURDFKHG� DW� D� VKRUW�GLVWDQFH� RI� WKH� PLQH�� 7KH� ZLUH� ZLOO� EH� FXW� XVLQJ� D�FXWWLQJ�GHYLFH�LQWHJUDWHG�LQ�WKH�KDQG��7KH�SDUW�RI�WKH�ZLUH�QH[W�WR�WKH�PLQH�ZLOO�EH�KROG�E\�WKH�KDQG��)LJ����DQG� WLHG� DURXQG� WKH� IXVH� WR� VHFXUH� WKH� VWULNHU�UHWDLQLQJ� SLQ�� $IWHU� WKDW� WKH�PLQH� ZLOO� EH� H[FDYDWHG�E\� WKH� URERW�KDQG�KROGLQJ� D� VKRYHO�OLNH� WRRO� DQG�E\�KLJK�SUHVVXUH� DLU�� 7KH� PLQH� W\SH� FODVVLILFDWLRQ� ZLOO�EH�SURYHG�DQG�WKH�JUDVS�SODQQLQJ�IROORZV��7KH�PLQH�ZLOO�EH�SXW�DVLGH�DQG�PDUNHG�ZLWK�D�FRORU�� �

)LJ���0LQH�FOHDUDQFH�KDQG�KROGLQJ�WKH�WULSZLUH�RI�D�320=���IUDJPHQWDWLRQ�$3�PLQH�

,Q� FDVHV� ZKHQ� D� YLVXDO� GHWHFWLRQ� GRHVQ¶W� ILQG� DQ\�REVWDFOHV� RU� PLQHV� WKH� PLQH� GHWHFWLRQ� VHQVRU� ZLOO�VFDQ�WKH�JURXQG��,I�D�GDQJHURXV�VSRW�ZLOO�EH�GHWHFWHG�ILUVW�� WKH� URXJK� FODVVLILFDWLRQ� RI� WKH� PLQH� W\SH� ZLOO�WDNH� SODFH�� 8VLQJ� WKH� GLVWDQFH� WR� WKH� EXULHG� REMHFW�DQG�LWV�VL]H�LW�FDQ�EH�JXHVVHG�LI�WKH�REMHFW�LV�DQ�$3�RU�DQ�$7�PLQH��,Q�FDVH�RI�DQ�$3�PLQH�WKH�VSRW�ZLOO�EH�PDUNHG�DQG�WKH�IROORZLQJ�PLQH�FOHDUDQFH�YHKLFOH�ZLOO�UHPRYH� WKH�PLQH�� ,I� LW� FDQQRW�EH�PDGH� VXUH� WKDW� WKH�REMHFW�LV�DQ�$3�PLQH�RU�LI�WKH�ORFDWLRQ�RI�WKH�PLQH�LV�QRW�UHDFKDEOH�E\�WKH�PLQH�FOHDUDQFH�YHKLFOH�� WKH�RQH�RI� WKH� WZR� URERWV�ZLOO� H[FDYDWH� WKH�PLQH��7KH� KDQG�ZLOO�KDYH� WR�JUDVS�D� WRRO� ILUVW� DQG�GLJ� WKH�JURXQG�RU�XVLQJ� WKH� KLJK�SUHVVXUH� DLU� RU� KLJK�SUHVVXUH� ZDWHU�IUHH� WKH� REMHFW¶V� VXUIDFH�� 8VLQJ� WKH� PLQH� GHWHFWLRQ�VHQVRU� WKDW� LQFOXGHV� D� *35� �JURXQG� SHQHWUDWLQJ�UDGDU�� WKH� GLVWDQFH� WR� WKH� EXULHG� REMHFW� ZLOO� EH�PHDVXUHG�DW�FHUWDLQ�WLPH�SRLQWV��7KH�OD\HU�RI�WKH�VRLO�ZLOO� EH� UHPRYHG� E\� KLJK�SUHVVXUH� DLU� LI� SRVVLEOH� WR�DYRLG� SUHVVLQJ� WKH� PLQH� VXUIDFH�� 2QFH� WKH� PLQH�VXUIDFH� LV� YLVLEOH� WKH� H[DFW� PLQH� W\SH� FODVVLILFDWLRQ�FDQ�EH�FDUULHG�RXW��7KH�JUDVSLQJ�RI�D�PLQH�LV�D�FUXFLDO�WDVN� EHFDXVH� D� WRR� VWURQJ� SUHVVXUH� RQ� WKH� PLQH�VXUIDFH�RU�WKH�VOLSSLQJ�RXW�RI� WKH�JUDVS�FDQ�FDXVH�DQ�H[SORVLRQ�� 7KH� SDUDPHWHUV� VWRUHG� LQ� WKH� PLQH� W\SH�GDWDEDVH� FDQ� KHOS� LQ� WKH� JUDVS� SODQQLQJ�� 7KH�

SUREDELOLW\� RI� VOLSSLQJ� RI� WKH� PLQH� IURP� WKH� VHFXUH�JUDVS� FDQ� EH� UHGXFHG�� 7KH� PLQH� FDQ� EH� JUDVSHG��H[FDYDWHG��SXW�DVLGH�DQG�PDUNHG�ZLWK�D�FRORU��

,Q�FDVH�D�PLQH�RU�DQ�8;2�WKDW�WKH�URERW�FDQQRW�GHDO�ZLWK��YHU\�ELJ�RU�KHDY\�RUGLQDQFH��ZLOO�EH�IRXQG�WKH�URERW�ZLOO�RQO\�PDUN�WKH�SODFH�ZLWK�D�FRORU�DQG�OHDYH�WKH�PLQH�RU�8;2�IRU�IDUWKHU�GLVDUPLQJ��

7RRO�FKDQJLQJ��7KH�FRQWURO�RI�WKH�URERW�PDQLSXODWRU�DQG�KDQG�ZKLOH�JULSSLQJ� WRROV� WDNHV� SODFH� DXWRQRPRXVO\� DIWHU�UHFHLYLQJ� D� FRPPDQG� IURP� WKH� RSHUDWRU�� 7KH�RSHUDWLRQ�RI� WRRO� FKDQJLQJ� LV� SUHSURJUDPPHG��7RROV�DUH� ORFDWHG� RQ� WKH� VKHOI� LQ� WKH� IURQW� RI� WKH� URERW�EHORZ� WKH�PDQLSXODWRUV��(DFK� WRRO�KDV�D�JULS�ZLWK���KROHV�WKDW�PDNHV�LW�HDV\�IRU�WKH�URERW�KDQG�WR�KROG�LW�ZLWK���ILQJHUV��7KH�WRROV�DUH��D�VKRYHO��D�JUDVV�FXWWHU��D�FKDLQ�VDZ�DQG�D�SURGGHU�WR�LQYHVWLJDWH�WKH�VWLIIQHVV�RI�WKH�JURXQG��� �����&RQWURO� 6WUDWHJ\�������0DVWHU�6ODYH�7KH� KDQG� DQG� PDQLSXODWRU� DUH� SODQQHG� WR� EH� WHOH�RSHUDWHG��7KH�PDVWHU�DQG�VODYH�PHFKDQLVP�ZLOO�DOORZ�D� KXPDQ� WR� FRQWURO� WKH�PRYHPHQWV� RI� WKH� KDQG� DQG�PDQLSXODWRU� LI� D� IXOO\� DXWRQRPRXV� FRQWURO� LV� QRW�SRVVLEOH�� 7KH� PDVWHU� LV� VHQGLQJ� FRPPDQGV� WR� WKH�VODYH� DQG� VRPH� LQIRUPDWLRQ� LV� IHG� EDFN� IURP� WKH�VODYH� WR� WKH� PDVWHU�� 7KLV� LV� WKH� VR�FDOOHG� ELODWHUDO�WHOH�RSHUDWLRQ�� �

�����)RUFH�VHQVRU�OHVV�KDQG�$V�D�SDUW�RI� WKH�PLQH�FOHDUDQFH� URERW�KDQG� UHVHDUFK�SURMHFW� D� IXQGDPHQWDO� VWXG\� RQ� WKH� UHGXFWLRQ� RI� WKH�QXPEHU�RI�VHQVRUV�LV�FDUULHG�RXW�� �

,Q� RUGHU� WR� LPSOHPHQW� JUDVSLQJ�� IRUFH� FRQWURO� LV�QHFHVVDU\�� ,W� LV� FRPPRQ� WR�XVH� IRUFH� VHQVRUV�RQ� WKH�ILQJHU� WLSV� WR� SURYLGH� SUHFLVLRQ� JUDVS� DQG� WRXFK�VHQVRUV� RQ� WKH� LQQHU� DUHDV� RU� WKH� KDQG� WR� SURYLGH�SRZHU�JUDVS��6LQFH�IRUFH�VHQVRUV�DUH�TXLWH�H[SHQVLYH�DQG�WKH�PHDVXUHG�GDWD�FRQWDLQV�D�ORW�RI�QRLVH�WKH�LGHD�RI� XVLQJ� IRUFH�VHQVRU�OHVV� KDQG� ZDV� ERUQ�� 8VLQJ�WRXFK� VHQVRUV� RQ� WKH� KDQG� LQQHU� DUHDV� GRHV� QRW�SUHYHQW� WKH�SRVVLELOLW\� WKDW�GXULQJ�JUDVSLQJ� WKH�KDQG�FDQ�EH�EORFNHG�E\� WKH�REMHFW�DQG� WKH�VHQVRUV�FDQQRW�GHWHFW�VXFK�D�EORFNLQJ��8VLQJ�WKH�MRLQW�PRWRU�FXUUHQW�DV� LQGLFDWRU� IRU� WKH� WRUTXH� PDNHV� VXUH� WKDW� DQ\�EORFNLQJ� RI� WKH� KDQG� E\� DQ� REMHFW� FDQ� EH� GHWHFWHG��7KH�DEVHQFH�RI�WKH�IRUFH�VHQVRUV�UHTXLUHV�KRZHYHU�DQ�DSSURSULDWH� KDQG� GHVLJQ� DQG� FRQWURO�� $Q� DGHTXDWH�KDQG� GHVLJQ� FDQ� VLPSOLI\� WKH� FRQWURO� V\VWHP�� )RU�H[DPSOH� D� KDQG� ILQJHU� ZLWK� WKUHH� URWDWLRQDO� MRLQWV�ZKLFK� D[HV� DUH� SHUSHQGLFXODU� WR� HDFK� RWKHU� DOORZV�IRUFH� GHFRXSOHG� PHDVXUHPHQW� LQ� DOO� WKUHH� �[�\�]��GLUHFWLRQV�RI�WKH���GLPHQVLRQDO�VSDFH��$UH�WZR�MRLQWV�URWDWLQJ� DURXQG� SDUDOOHO� D[HV� WKHQ� WKH� H[WUDFWLRQ� RI�WKH�IRUFH�LQIRUPDWLRQ�LV�PRUH�FRPSOH[��

�����+DQG�FRQWURO�FRQFHSW�7KH�PRYHPHQW�RI�WKH�URERW�KDQG�DQG�LWV�ILQJHUV�ZLOO�EH� FRQWUROOHG� DXWRQRPRXVO\� LQ� D� SUHSURJUDPPHG�PRGH��LQ�FDVH�RI�NQRZQ�REMHFWV��RU�WHOH�RSHUDWHG�E\�KXPDQ�KDQG�XVLQJ�D�GDWD�JORYH�DQG�D�PDVWHU�GHYLFH�IRU�IRUFH�IHHGEDFN��LQ�FDVH�RI�XQNQRZQ�REMHFWV���7KH�PRYHPHQW� RI� WKH� GDWD�JORYH� GHILQHG� E\� WKH� JUDVS�SDUDPHWHUV� >S � �S � �«�S � @� ZLOO� EH� FRPSDUHG� ZLWK� WKH�GDWD�EDVH�JUDVS� W\SHV��7KH�GDWDEDVH�FRQWDLQV� OLPLWHG�QXPEHU� RI� JUDVS� W\SHV� LQ� RUGHU� WR� OLPLW� WKH�LQIRUPDWLRQ� IORZ� IURP� WKH� PDVWHU� WR� WKH� VODYH�� 7KH�JUDVSV�VSHFLILHG�LQ�WKH�GDWDEDVH�DUH�FLUFXODU��VSKHULFDO�DQG�RWKHUV��$IWHU�WKH�DSSURSULDWH�JUDVS�W\SH�LV�FKRVHQ�WKH� JUDVS� VL]H� DQG� WKH� JUDVSLQJ� IRUFH� ZLOO� EH�HVWLPDWHG� IURP� WKH� JUDVS� SDUDPHWHUV�� 7KH� JUDVS� RI�WKH�VODYH�* � �LV�WKHQ�GHILQHG�E\�WKH�JUDVS�W\SH�* � � � ��� ���WKH�JUDVS�VL]H�* � � � � � �DQG�WKH�JUDVSLQJ�IRUFH�* � � � ��

=

��

� � ����

� ��� ��

�

***

*�

�

�

�

ZKHUH� WKH� JUDVS� W\SH� * � � � ��� � LV� D� GDWD� EDVH� IXQFWLRQ�'%�RI�WKH�JUDVS�SDUDPHWHUV��

( )�� ����� SSS'%* ������ !"# = �

DQG� WKH�JUDVS�VL]H�* � � � � � �DQG�JUDVS�IRUFH�DUH�GHILQHG�E\�WKH�DSSUR[LPDWLRQ�IXQFWLRQ�$�DQG�%��

( )$% & '�(% SSS$* ������ )*+ =( )$,% SSS%* ������ )*+ = �

7KH�HUURU�EHWZHHQ� WKH� VODYH�JUDVS�* � � DQG� WKH�REMHFW�VKDSH� DQG� VL]H� ZLOO� EH� PLQLPL]HG� XVLQJ� VHQVRU\�LQIRUPDWLRQ�DV�WKH�KDQG�LV�DSSURDFKLQJ�WKH�REMHFW��

����&RQFOXVLRQ��,Q� WKLV� DUWLFOH�ZH�SURSRVHG� WKH�PLQH� FOHDUDQFH�KDQG�ZKLFK� ZLOO� EH� FRQWUROOHG� PDLQO\� E\� WHOH�RSHUDWLRQ��7KLV�FRQFHSW�LV�XVLQJ�D�PLQH�FOHDUDQFH�KDQG�PRXQWHG�RQ� D� VL[�OHJJHG� PLQH� GHWHFWLRQ� URERW�� 7KH� PLQH�FOHDUDQFH� PHWKRGV� DQG� WKH� FRQWURO� VWUDWHJ\� ZHUH�EULHIO\� SUHVHQWHG�� $OVR� D� IRUFH�VHQVRU�OHVV� KDQG�GHVLJQ�ZDV�SURSRVHG�LQ�RUGHU�WR�UHGXFH�WKH�QXPEHU�RI�VHQVRUV�� ,Q� SDUWLFXODU�� ZH� WKLQN� WKDW� PRVW� DQWL�SHUVRQQHO� PLQHV� ZLOO� EH� FOHDUHG� E\� DQ� XQPDQQHG�YHKLFOH�� +RZHYHU�� VRPH� DQWL�SHUVRQQHO� PLQHV� OLNH�WULS�ZLUH�PLQHV� VKRXOG�EH�GLVSRVHG�E\� WKH�SURSRVHG�KDQG�� $OVR�� DQWL�WDQN� PLQHV�� VRPH� NLQGV� RI� VPDUW�PLQHV�DQG�8;2V�VKRXOG�EH�WUHDWHG�E\� WKH�GH[WHURXV�KDQG� EHFDXVH� RI� WKHLU� H[WUHPHO\� KLJK� GHVWUXFWLYH�SRWHQWLDO� LQ� FDVH� RI� DQ� H[SORVLRQ��:H� KRSH� WKDW� WKH�

SURSRVHG� KDQG� ZLOO� EHFRPH� D� YHU\� XVHIXO� WRRO� IRU�GHPLQLQJ�RQ�DQ�DFWXDO�PLQH�ILHOG�VRRQ��

����5HIHUHQFHV��>�@�$�%LFFKL�´+DQGV�IRU�'H[WHURXV�0DQLSXODWLRQ�DQG�

5REXVW� *UDVS�� $� 'LIILFXOW� 5RDG� 7RZDUGV�6LPSOLFLW\´�� ,(((�7UDQVDFWLRQV�RQ�5RERWLFV�DQG�$XWRPDWLRQ��

>�@�'DYLG� 0�� /DQH�� -�� %UXVFH� &�� 'DYLHV�� *UDKDP�5RELQVRQ�� 'HVPRQG� -�� 2¶%ULHQ�� -LP� 6QHGGRQ��(XDQ� 6HDWRQ�� $QGHUV� (OIVWURP�� ³$0$'(86�'H[WHURXV� 6XE� VHD� +DQG´�� ,(((� -RXUQDO� RI�2FHDQLF� (QJLQHHULQJ�� 9RO����� 1R���� -DQXDU\��������

>�@�7�5DSDUHOOL�� *�0DWWLD]]R�� 6�0DXUR��0�9HODUGRFFKLD�� ³'HVLJQ� DQG�'HYHORSPHQW� RI� D�3QHXPDWLF� $QWKURSRPRUSKLF� +DQG´�� -RXUQDO� RI�5RERWLF�6\VWHPV��������������������

>�@�<�,VKLNDZD�� <�:HQZHL�� +�<RNRL�� <�.DND]X��³'HYHORSPHQW� RI� 5RERW� +DQGV� ZLWK� DQ�$GMXVWDEOH�3RZHU�7UDQVPLWWLQJ�0HFKDQLVP´�3URF��RI�WKH��� - . �$QQXDO�&RQI��RI�WKH�5RERWLFV�6RFLHW\�RI�-DSDQ�6HSWHPEHU��������

>�@�+�6XJLXFKL�� 6�:DWDQDEH�� 5�� )XNDQR�� 7�0RULQR��+�<DJL� ³$� GHYHORSPHQW� RI� WKH� ILYH� ILQJHU� URERW�KDQG�ZLWK�GLVWULEXWHG�WRXFK�VHQVRUV´�3URF��RI�WKH��� - . � $QQXDO� &RQI�� RI� WKH� 5RERWLFV� 6RFLHW\� RI�-DSDQ�6HSWHPEHU��������

>�@�&�0�6HJXQD�� ³7KH� 'HVLJQ�� &RQVWUXFWLRQ� DQG�7HVWLQJ� RI� D� 'H[WHURXV� 5RERWLF� (QG� (IIHFWRU´��8QLYHUVLW\� RI� 0DOWD� 5�<RVKLQR�� ³+DQG� RI�+XPDQRLG� 5RERW´�� 5RERWLFV� 6RFLHW\� RI� -DSDQ��9RO�����1R�����������

>�@�7K�)LVFKHU�� -�6H\IULHG�� ³7KH� QHZ� .DUOVUXKH�'H[WHURXV�+DQG�,,´��8QLYHUVLW\�RI�.DUOVUXKH�

>�@�0�5�&XWNRVN\��³2Q�*UDVS�&KRLFH��*UDVS�0RGHOV��DQG� WKH� 'HVLJQ� RI� +DQG� IRU� 0DQXIDFWXULQJ�7DVNV´�� ,(((� 7UDQVDFWLRQ� RQ� 5RERWLFV� DQG�$XWRPDWLRQ��9RO���1R����-XQH��������

>�@�86$� 'HSDUWPHQW� RI� 'HIHQVH� ³7HFKQRORJLHV� WR�$VVLVW� ZLWK� WKH� *OREDO� 'HPLQLQJ� (IIRUW´�� VKRUW�PRYLH�)HEUXDU\��������

A “TOOL KIT” FOR DEMINING ROBOTS

P. Kopacek

Institute of Handling Devices and Robotics Vienna University of Technology

Favoritenstrasse 9-11, A- 1040 Vienna, Austria kopacek@ihrt.tuwien.ac.at

Abstract: A tool kit for demining robots will be presented. It consists of several, mostly commercially available mobile, intelligent platforms and various peripheral devices. All the modules are compatible in hard- and software. This tool kit offers the possibility to assemble in a very easy way robots for different deminig tasks. Keywords: Robots, demining, robot swarms, modular robots. 1. Introduction Robot will be more and more used for demining tasks. There do not exist standard equipment requirements for the humanitarian demining. But the following requirements for systems, that should improve the work of demining, can be summarized [3]:

• The system (robot, vehicle or equipment) should be “low-cost” affordable to countries not able to use the expensive systems available on the market.

• Components should be commercially available in order to reduce maintenance costs and availability problems.

• The system has to be “transportable”. • Robust design for working in rugged

environments is necessary. • User-friendly “HMI – Human-Machine

Interface” reduces training problems. • Operation time should be at least 2 hours.

Re-fueling or re-charging should not take more than 30 minutes.

• Ability to distinguish mines from false alarms like soil clumps, rocks, bottles and tree roots.

• Operation in a variety of soil types, moisture contents and compaction states.

• Ability to detect all types of mine of different types and sizes.

• Operation in vegetated ground cover. These features require new “single purpose” robots.

The beginning of multipurpose robots was characterized by the effort to design robots, which can be used for many different tasks and applications. The idea was that these universal robots would be produced in such large numbers that the price is reasonably low. One solution could be a modular system – a “tool kit” – for mobile robots. The general philosophy of modularization of mobile robots is that a manufacturer produces a number of modules by which user can assemble such configurations of robots which are the most appropriate for a specific task. 2. Mobile, intelligent robot platforms Mobile Robot Platform The basis of a modular concept for demining robots is the Mobile Robot Platform (MRP) which can be described as a multi-use mobile robot, developed in its basic configuration and having all the most important and vital for its mobile functions systems. These platforms can be divided in some basic systems:

• Locomotion system • Driving system • Main control system • Communication system

Locomotion System The locomotion system can be realized on different principles, like wheel-, chain-, walking- or special-locomotion. All these principles have different characteristics in regard to costs, weight and efficiency in various terrains. Driving System The driving system consists of: power sources, actuators and transmissions, which enable the platform to perform the necessary movements.

Main Control System The main control system is usually microprocessor based and responsible for the actions executed by the robot. It should be powerful enough to meet the requirements of all current tasks and possible future tasks. Navigation System The navigation system normally consists of an array of, possibly different, sensors for the perception of the robots environment and of course of some kind of software which designates the rules for the robot-movement. Communication System The communication system connects all the systems in the platform and the platform with the environment. 3. A robot “tool kit” “Modular Mobile Robot is intelligent (low, medium or high degree of intelligence) semi or fully autonomous vehicle (wheel, legged, chain, crawling, climbing or special locomotion) with all its systems (locomotion, driving, control, navigation and communication) build on a modular principle, able to carry peripheral systems (robot arms etc.) or tools (conventional or special) for transporting of loads or executing different industrial or service operations in world coordinates.” [4] Arms and Peripherals with Less than 3 DOF The mobile robot platform has to be equipped by a mechanical or other system able to manipulate tools or grippers and operate with them. Simple arms, lifts,

fork mechanisms etc. are needed to perform some simple tasks. Though these tasks could be performed by a dexterous powerful arm as well, it would be simply too expensive. Arms with More than 3 DOF Some operations performed by mobile robots require functions of flexible robotic arms with higher number of degrees of freedom. Some of these arms are controlled by the main control system of the mobile robot platform others are controlled by a special microcontroller or by a Programmable Logic Controller (PLC). Some of these sophisticated systems installed on board of mobile robot platforms are independent robots having all-typical subsystems. Of course there is always some communication and interaction with the platform but theoretically such onboard robots can perform task autonomously. Grippers Grippers are designed to imitate the human hand and operations which are fulfilled by it. It is highly complicated to imitate the complex motion sequences of a human hand. Therefore most of the grippers are only a simplified copy with less DOF. There are different categories of grippers used in robotics. Mechanical grippers apply certain forces by fingers as to keep the object safely grasped. They can be actuated by all common principles and are often equipped with sensors to avoid damaging the handled object. Vacuum grippers are mostly applied for sensitive and fragile materials. The surface characteristic of the handled object limits the applicableness of vacuum grippers. Electromagnetic grippers are used for manipulating magnetic sensitive materials. There is of course always the possibility to use special grippers designed for specific tasks. Tools The mobile robot platform can be upgraded and modified by adding a number of peripheral systems and tools for the performance of different tasks or functions. There is a large variety of tools, which can be used. Basically these tools can be divided into two major categories:

Conventional tools Special tools

Sensor

User panel

Control Computer

Power Supply

Ultrasonic/ Infrared

Passive wheel Active wheel

Steering/ Drive

System

Fig.1 Mobile robot

Conventional Tools Conventional tools are similar in regard to their function to conventional hand-held tools for manual operations. The difference is in their design, since they have to be fixed on the mobile robot platform, and actuation. Different actuating systems for the tools are applied:

• Electromechanical • Electromagnetic • Pneumatic • Hydraulic

Examples for conventional tools:

− Screw drivers − Drilling tools − Polishing tools

Most conventional tools are attached to an onboard robotic arm and controlled via a special microprocessor system and a number of sensors which enable the performance of a variety of operations automatically. Special Tools A special tool installed onboard of a mobile robot platform changes the same to a specialized mobile robot system. When special tools are lightweight constructed the manipulation system can be more flexible and with wider reach. Heavier tools cannot be very flexible. They need more rigid and strong manipulation systems. So there is often only one degree of freedom applied, and the other DOFs are realized by the mobility of the platform. The variety of tasks realized by service robots and especially personal robots is much broader than of those in the production sector. This is why they require a much larger number of variations of the robot peripherals. Tool changing system Installing a tool changing system enables the robot to achieve a wide variety of performable operations. Tool changing systems are normally placed at the end of a robot arm. They have to be light, simple and very reliable. They use again electrical, electromagnetic, pneumatic and hydraulic actuators. As they work with tools of different actuation principles they use the same type of actuation as the tools themselves. There have to be also onboard magazines which host the tools. These magazines have to ensure easy access and must be able to free the used tool and output another easily and safely.

Additional Sensors The basic configuration of each mobile robot platform has its integrated sensors. The navigation system makes excessive use of sensor for position determination and collision avoidance. But there are numerous possibilities to upgrade the system with additional sensors for some special applications or to extend its abilities. Storage Devices In many mobile robot applications transportation is an important part of the overall task. To transport different items mobile robot platforms have to be upgraded with another type of peripheral devices: special storage systems or devices. Storage devices have to be designed with regard to the required space, loading and unloading conditions and possible special environmental conditions for the transportation of sensible items. Interfaces The crucial point to make a modular system work is an appropriate connection between the mobile robot platform and the modules and among different modules themselves. The hardware interface should ensure a proper mechanical connection and concurrently ensure power and information supply. Software interfaces are at least as important as hardware interfaces. Without a good cooperation in the software layer new added modules can not complete their tasks and are therefore of no use.

Mobile Platform

Interfaces

Hardware Software

Storage devices

(Transportation of mines )

Robots > 3 DOF.

Arms < 3 DOF.e.g. lifts

g r I p p e r s

CommunicationModule

MAS Docking station

Sensors

Tool changingsystem

ControlPLC

Special tools

Removing

Conv.tools tools

Detection Navigation

Figure 2: Modular Robot System [4]

Communication Modules MAS Although mobile robot platforms are normally equipped with a communication system it could be necessary to use some special communication systems. Especially in multi agent systems (MAS) where a team of robots acts together is communication between the team members of importance. Docking Stations There are two types of docking stations

Battery recharging (or fuel, compressed air filling) docking stations

Loading unloading stations Batteries are today the most used energy sources for mobile robots. Combustion engines are used for heavy mobile robots working open-air. There are worldwide efforts to develop new or more efficient energy sources. An automatic recharging system for mobile robots would be desirable as to be able to operate continuously. This would include sensors to notice the necessity of recharging the system. Then the mobile robot has to move to the docking station and refuel. Good designed systems should assure an appropriate connection of the robot to the docking station, an optimal recharging process and a proper fulfillment in regard to the accurate amount.

4. A ‘Tool Kit’ for Demining Robots The former chapter dealt with a general modular concept for mobile robots. Here the application for demining actions will be discussed.. A tool kit consists of two or three different mobile platforms and a number of peripheral devices . Assuming a mine infected area has been made out the subsequently clearance process can be divided into following steps. 1. The exact positions of each single mine should be

known. Therefore the mine has to be detected by some sort of detection system and the position of the mine should be stored someway.

2. The mine has to be removed which could mean in the case of a buried mine that it has to be excavated. It might be that the mine gets defused before removing it completely. But there are landmines in use which can not be deactivated. Another possibility would be to detonate the mine by using an explosive charge. Since mines include a lot of chemicals which get into the ground in case of a detonation it is better not to detonate them. Mine cleared areas are often used for farming, hence a chemical pollution of the soil is undesirable.

3. If the mine was not detonated it has to be transported out of the area and collected at some place.

Considering these actions it is obvious that one single robot designed to accomplish all these tasks has to be a pretty sophisticated device. On the other hand it can

Mobile Platforms

Wheeled Chains

Legs

Communication LAN

Host Other robots

OwnSwarm Other Swarm

Robot arm Gripper

Storage

Sensors

Detection Navigation

Metall Sniff GPS Laser Others

Docking Station

Docking Station

ScrewDriver others

PLC Human MachineInterface

Fig. 3 “Tool Kit” for demining robots

be seen that there could be employed at least two different types of robots. One type is responsible for detecting and one for removing the mines. It is likely to use a third robot-type for transporting the removed mines out of the area. Another advantage is the flexibility. For example it could happen that the terrain of the minefield changes dramatically and another type of locomotion system would be now the system of choice. Using a modular robot system, it would be easy to change the locomotion module or the platform and continue the work. Different types of soil or different types of landmines may demand different removal tools. No problem for a modular system. Each robot used for demining has to be assembled by using the modules of the ‘Tool Kit’. The goal of inventory design is to create the smallest inventory of modules that can be assembled into the largest diversity of robots. Although the task of humanitarian demining is clearly defined (i.e. detection, removal and transportation). Unforeseen problems should be solved with available resources one of the design specifications should be adaptability. Therefore in inventory design, the level of modularity is important. A low-level inventory would contain very basic elements such motors, gears, bearings and nuts and bolts. A high-level inventory would contain complex elements such as limbs or arms. A low-level inventory offers more flexibility in the variety of robots to assemble. But the assembling is more sophisticated. Conversely a high-level inventory allows fewer robots but the assembly is simplified. When it comes to the realization of such a tool kit many different factors have to be minded and many decisions have to be made before even starting with the design. The decision about the level of modularity will depend on cost factors, technical feasibility, variety of performable tasks and many other things. The highest level of modularization would be the use of robots made from identical modules. These robots are in a prime state of development and their abilities have only been demonstrated to a less degree. But apart from this it is absolutely imaginable to include in a tool kit translatorial and rotational modules which can be assembled to different robot structures (e.g. articulated, Cartesian, spherical, cylindrical, SCARA, …) which can be attached to the mobile robot platform. Such modules like in Fig. 3 are normally composed of a motor, precision reduction gear, position encoder, brake, motor amplifier, limit switches, on-

board computer controller, and internal cabling. The fabrication of such modules should not be, in case of a reasonably number of pieces, too costly. This is quite important because the robot arm for removal actions is likely to be one of the most endangered parts to be damaged by an explosion. Beside these actuation modules there is also need for kinematic modules. They are used to alter the dimension of the robot which means to change the distance between the robots joints. This greatly affects the capability of the robot in terms of strength, reach and accuracy. If the tool kit consists of several actuation modules of different sizes there are also adapter modules required to link modules with differential interface sizes. To build functional robots each module must be capable of interfacing with all other modules. The interface can be broken into three categories:

• Mechanical Interface • Power Interface • Information Interface

It is necessary that the modules can be easily connected with each other to make the robots reconfigurable. This quality of the modules also reduces the maintenance effort and simplifies the exchange of defect modules. A quick-coupling mechanism with which a secure mechanical connection between modules can be achieved would be an excellent solution. At the same time the mechanical connection is made, power and information connection should also be made. There are many possible solutions for this problem. The following picture shows one of them. The transmission of power between modules depends on the power source. If pneumatic and/or hydraulic actuators are used an electrical bus for transmitting electrical power will not be sufficient. Especially for heavier tasks like the excavation of landmines hydraulic actuators may used. Considering for example a simple translatorial or rotational module used for a robot arm it could be necessary only to link up the end effector with the hydraulic powering aggregate. The necessary information transfer between modules can be done using electrical or optical connections. Information transfer can occur in many ways. Each module will need its own processor to handle communication between modules and local control (e.g. position joint control). 5. Summary Robots for demining are today mostly “single purpose” devices. Therefore these robots are expensive not only because of the small lot sizes. A

solution could be the general modular concept for mobile robots described in this paper. Such a concept of a modular robot system in form of a tool kit can be easily adapted for demining actions. A tool kit can be imagined as a package of modules like that one discussed before. It may consist of two or three different mobile platforms and an indefinable number of peripheral devices to assemble different configurations according to the expected tasks. 6. Literature [1] Baudoin, Y. et.al. (2000): “Humanitarian

Demining: Sensory and Robotics”. Proceedings of the IMEKO World Congress 2000, Vienna, Vol. XI, p. 241 – 251.

[2] Bunzl, R.: “A toolkit for demining robots”. Diploma Thesis, Vienna University of Technology, 2002 (in print).

[3] Kopacek, P. (2002): “Deminig Robots – a tool for International Stability”. Proceedings of the 16th IFAC World Congress, Barcelona, July 2002.

[4] Shivarov, N. (2001): “A tool kit for modular, intelligent, mobile robots”. PhD. Thesis, Vienna University of Technology, 2001.

[5] Shivarov, N: “Modular, intelligent, mobile robots”. Proc. of the “1st Bulgarian – Austrian Automation Day”, Sofia, 2001 (in print).