robotic assembly - springer · assembling experience tells that adding hopping motion along the...

Robotic Assembly

Heping Chena*, Biao Zhangb and George ZhangbaIngram School of Engineering, Texas State University, San Marcos, TX, USAbABB Corporate Research Center, ABB Inc, Windsor, CT, USA

Abstract

At the present time, industrial robots for assembly tasks only constitute a small portion of the annualrobot sales. One of the main reasons is that it is difficult for conventional industrial robots to adapt tothe complicity and flexibility of assembly manufacturing processes. Therefore, intelligent industrialrobotic systems are attracting more and more attention. This chapter discusses robotic assemblytechniques that perform assembly tasks with part geometric variations, part location variations and/or fixture errors. Different assembly tasks were implemented to demonstrate different techniques.For complex assembly processes, assembly parameters are very critical for assembly cycle time andFirst Time Through (FTT) rate. Hence the exploration of optimal parameters to minimize the cycletime and maximize the FTT rate has to be discussed. The Design-of-Experiment (DOE) method isadopted to identify the optimal parameters and experimental results demonstrate the effectiveness ofthe proposed DOEmethod. Since the proposed techniques were tested using real industrial assemblyprocesses, they are ready for industrial implementation.

Introduction

Assembly tasks using industrial robots have increased in both number and complexity over the yearsbecause of the increasing requirements of product quality and quantity. However, assembly robotsare still a small portion of total robot sales each year. One of the main reasons is that it is difficult forconventional industrial robots to adjust to any sort of change. Therefore, more intelligent industrialrobotic systems are rapidly expanding the realms of possibility in assembly applications becausethey can perform assembly tasks with high autonomy and adaptability to the environments. In thischapter, several robotic assembly techniques are discussed for different industrial applications, suchas assembly without high precision fixtures, assembly on a moving production line, and highprecision assembly.

Robotic force control strategy is developed for high precision assembly while controlling thecontact force/torque. Since the force/torque control is sensitive to contact with the environment, thetool position/orientation can be accurately controlled. This makes the force control assembly a bettersolution for robotic assembly. There are many successful stories using force control assembly inindustrial applications.

*Email: [email protected]

Handbook of Manufacturing Engineering and TechnologyDOI 10.1007/978-1-4471-4976-7_105-1# Springer-Verlag London 2014

of 47

However, the force control method requires extra devices, such as the force/torque sensors andcontrol software package, etc., which makes the robot control system more complicated and moreexpensive. Therefore, a method to perform assembly process without force control is also discussedin this chapter. Because the force control hardware and software are eliminated, this implementationof such technique is much simpler.

The motion range of force control has certain limitations because it is difficult to keep the roboticsystem stable and sensitive within a large range of force-controlled motion. Typically vision systemis installed to compensate for large part location error. Force control is then applied for fine assemblyposition searching. In this case, the requirements for both force control and vision system arereleased.

Because many production lines are moving, especially in automotive manufacturing, assembly ona moving production line is worth discussing. The natural frequency of industrial robots is typicallylow, while the moving production lines contain both high-frequency and low-frequency terms.Therefore, visual servoing and force control are integrated to control the motion of a robot to trackthe motion of a moving production line while assembly tasks are performed.

Robot force control introduced complexity and uncertainty to the robot programming, controlparameter setting up, and manufacturing process quality. The force-controlled robot behavesdifferently for different contact force conditions resulting from the manufacturing variations of theassembled parts, fixture, and environment disturbances on the manufacturing floor. One of the mostrecognizable behavior differences from position-controlled robot is that the production cycle time isno longer a predetermined value for force-controlled robot. Assembly parameters have large impacton the cycle time and FTT rate. Therefore, they have to be optimized in order to achieve optimalassembly performance.

As aforementioned, these topics will be discussed in the following sections. Experimentalsamples are used to demonstrate the techniques.

Robotic Assembly Techniques

Force Control AssemblyIntroductionTraditionally, industrial robots are used for painting, material handling, welding, and othernoncontact manufacturing processes. Recent advancement in robot control systems permits anentire new class of robot behaviors and applications. The new robot behaviors are possible due tothe incorporation of force feedback into the robot control system. Industrial robot applications cannow be expanded into processes with contact forces such as assembly and machining because theforce of contact with the environment in these applications can now be controlled.

Adding force feedback to the robot force control system allows dynamic alterations to the robottrajectory to control the contact force. Sicilinano et al. (1999) give basic robot force control concept,theory, and implantation. The force control technology enables robotic automation applications thatmate parts together such as gear meshing, spline insertion, clutch hub assembly, and surface grindingfollowing complex curved geometry with simple programming constructs. Zhang et al. (2004)reveal a typical robot force control assembly system for articulated robots. Around the same time,attempt has been done to develop force-controlled parallel robot structure to deal with the assemblyrequirement for higher sensitivity. Morris et al. (2001) describe an example of such system. Based onthe force control feature developed, many assembly applications have been developed for automo-tive power train assemblies such as torque converter assembly (Morris et al. 2001) and engine piston


of 47

installation (Chen et al. 2007). While the principles are the same, there are different techniques andcharacteristics for each application. This section will briefly describe the robotic force controlalgorithm and implementation and the assembly application examples, and finally a summary willbe provided to address the result, major lessons learned, and further work or the trend of the roboticassembly technology and application.

Force Control Theory and DevelopmentForce control is an enabler for industrial robotics to be used in sophisticated assembly applications.While industrial robots are traditionally used for painting, material handling, welding, and othernoncontact manufacturing processes, force-controlled robots can handle contact operations such asassembly and machining. Normally, robotic force control employs a 6D transducer or called forcesensor, being mounted on the robot faceplate. In some other cases, torque sensors are placed on eachjoint of the parallel or articulated robot arms to measure the joint torque values.

Most existing industrial force control is built on previous position control systems. An analog ordigital channel is added on the robot controller to take the force sensor feedback. The feedback thenintegrated into the robot control loop to realize the force control. Figure 1 is an example hardwaresetup for industrial robot force control.

Figure 2 shows a typical force control loop diagram. Reference force and reference velocity areinput into the control system as robot force control behavior desired by the user. Properly designedforce/velocity input can generate corresponding searching or other robot reaction patterns.

Because the robotic force control assembly requires a robot to perform assemblies in which theassembly tolerance is small, a search pattern has to be implemented in order to compensate for thepart positioning errors. A spiral search pattern is used in the X and Y searching directions.

x ¼ R sin a sin by ¼ R sin a cos b

(1)

The search pattern is shown in Fig. 3.When using force control in assembling, the searching pattern or searching strategy needs to be

defined first. Most time the strategy is coming out from the experience of manual assembling workand the part design analysis. Different force control developers may use different terminology, but it

Fig. 1 An example of hardware setup for industrial robot force control


of 47

involves search pattern such as linear search, circular search and spiral search illustrated in Fig. 3,insertion force, and search orientation arrangement. In most automotive part assembly, the orienta-tion is fixed or with a little flexibility, which is controlled by dumpling factor D as illustrated inFig. 2. In some of the search patterns, there are other related parameters such as search radius to bedefined and optimized to get out the optimal assembly system result for specific applications. Whileparameter optimization methodology and practice will be introduced in the later sections, the searchstrategy will be further described and explained in the next section (Experiment and Examples).

Experiment and ExamplesTorque Converter Assembly Automotive torque converter assembly is a typical robotic assemblyexample. Torque converter can be weighed up to 75 kg. Between torque converter and transmissionhouse, normally there are three distinct subassembly stages required to be matched during theassembly process. Figure 4 illustrates the gear matching for torque converter assembly. Assemblingtorque converter manually causes working injury and other ergonomic problems. Automating torqueconverter assembly is demanded by the automotive industry. Robotic force control provides a bettersolution than hard automation equipment with more flexibility and cost advantages. From Fig. 4, it

Fig. 3 A spiral search pattern to find the exact position for force control assembly. The units are mm

Fig. 2 A typical robot force control diagram


of 47

can be seen that the torque converter assembly is an insertion process with several layers of gearmatching. A typical torque converter assembly has three layers of matching: first layer is the first(largest) gear of the torque converter to the third gear (from top down) of the transmission shaft; thesecond matching is the second gear of torque converter to the second gear of the transmission shaft;and the third matching is the grove at the tip of the torque converter shaft to a bar on the bottom of thetransmission shaft. For the first layer matching, spiral search pattern is used; for the second layer,circular or spiral search pattern is used; and for the third layer, only rotation pattern is used becausethe assembling and assembled parts should be already lined up when reaching the third layer. Interms of rotation angle ranges, they are different for the three layers. First two layers use relativesmall angles such as 45� since the matching gears have fine teeth (as can be seen in Fig. 4). The thirdlayer uses larger searching rotation angle, at least more than 90�, since it has only two-gear teeth.Assembling experience tells that adding hopping motion along the axial direction will help improv-ing the assembly cycle time because the constant axial insertion force will press the two assemblingsurfaces together, which prevents gear matching during the assembly.

To summarize the above observation and analysis, for the robotic force control torque converterassembly process, the following force control assembly strategy is used:

• Use straight insertion and divide the assembly process into three or more layers.• Use spiral search pattern for the first layer, circular or spiral for the second layer, and no radial

direction search for the third layer.• In the first two layers, relative small rotational search angles, about 45� or less, are used; but in the

third layer, a large rotational search angle, more than 90�, is used.

Fig. 4 Gear matching between torque converter and transmission housing


of 47

• Hopping motion (in terms of insertion force vibration) is added along the axial direction toimprove the cycle time through overcoming gear surface “sticking” problem.

Figure 5 shows a robotic torque converter assembly along the horizontal direction. In this case, thegravity effect will not help in finalizing the assembly process (gravity will not hold the assembledpart in place along horizontal assembling as in vertical assembling). Therefore, a final positionlocking mechanism is needed to secure the torque converter in place when the assembly is finishedand robot releases the part. Other strategies for the vertical assembly process are applicable to thehorizontal assembly process.

Engine Piston Installation Another example of robotic force control assembly is engine pistoninstallation. Because the part tolerance and chamfer are getting smaller on the piston head for higherenergy efficiency, manually stuffing the piston is a challenge especially for diesel engines where thepiston is heavier, making manual operation even more difficult. Robotic force control pistoninstallation can combine the piston stuffing with cam bolt placement and tightening in one roboticautomated station, which increases the productivity and reduces the manufacturing floor spacerequirement. Figure 6 shows an example of engine piston stuffing in lab experiment. Multiple robotscan be used in the engine piston assembly. No matter how many robots are used, only one robotneeds to have the force control feature to perform the piston stuffing operation.

Fig. 5 Robotic torque converter assembly horizontally

Fig. 6 Engine piston stuffing/installation lab experiment


of 47

The gas engine piston stuffing/installation is one of the most challenging assembly processes.Because of the tight tolerance (up to 0.05 mm between the piston skirt and cylinder bore), thecomplexity of the ring quizzing, the necessary hairpin guidance, as well as the precision stuffingpushing, manual work is the most common practice, and hard automation is sometimes used.

In manual engine piston installation, two persons normally cooperate with each other in theassembly process. One person inserts guiding pins through a cylinder bore in the engine block tolocate and align the holes on the piston connecting rod, and another person holds the piston witha piston ring compressing device and approaches the cylinder bore to receive the guiding pins withthe piston connecting rod. A pneumatically driven pusher is often used for stuffing the piston into thecylinder bore. The connecting rod cap and screws are placed by hand, and then the cap is fastenedmanually using hand tools. Manual piston installation work is labor intensive, tedious, and prone toworker injury due to the force required and repetitive nature of the tasks. The assembly quality isentirely dependent upon the skill and attention of the workers.

Automation of engine piston insertion has been performed by using specially built machines,often called “hard automation.” These dedicated machines are huge, costly, and inflexible.Switching between engine models or types to be assembled is difficult, time consuming, and costly.Piston installation with a dedicated automation machine requires a level of precision and tolerancethat limits its application. Furthermore, in “hard automation” stuffing techniques described above,there is no active searching action in finding the cylinder bore. Though a passive floating table issometimes used to align the piston skirt with the cylinder bore, the success of stuffing the pistonessentially depends on the accuracy of the machine, the actual gap between the cylinder bore and thepiston skirt, and the leading chamfers on both the piston skirt and the cylinder bore. With theincreasing demand of reducing the gap between the cylinder bore and the piston skirt and minimiz-ing or eliminating the leading chamfers for the purpose of emission control and engine efficiencyimprovement, the challenge and difficulty are increasing for both manual and automated pistonstuffing processes.

To deal with this complicated engine piston installation process, an automated piston installationtechnology using industrial robots is developed. A three-robot and a two-robot workcell configura-tions are employed to balance the requirements of cycle time, floor occupation, and systemcomplexity. The piston installation process has been simulated; the special robot gripper for thepiston stuffing is designed and manufactured; robotic force control searching strategy is defined; andparameter setting and other programming-specific issues are investigated. After the robotic pistonstuffing shown in Fig. 6 has been successfully performed in the laboratory, a customer demo unit isbuilt as shown in Fig. 7. A three-robot piston installation system is described in this example. Other

Fig. 7 Robotic engine piston installation customer demo


of 47

cell configuration with fewer robots and more specific hardware can be also used to reduce thesystem complexity and cost.

The three-robot configuration is a total automation solution. As illustrated in Fig. 8, robot I with itsfixture picks up the engine block from a pallet and moves the engine block to a position close to thesecond robot and orients the engine block so that robots II and III can perform the piston installation.The fixture with engine block is shown in Fig. 9. An external axis (motor) controlled by the robotcontroller rotates the crankshaft of the engine block to the proper orientation for the piston to bestuffed in an associated one of the three cylinder bores. As shown in Fig. 10, the piston installationoperation consists of stuffing the piston with ring skirt, connecting rod assembly into one of thebores, and attaching the associated connecting rod cap to that assembly. Robot II which has forcecontrol moves its uniquely designed gripper. The gripper has jaws to grab the piston assembly.Before the jaws are closed to grab the piston assembly, the gripper sucks up the piston assemblyusing a suction cup on a pusher against the upper surface of the piston head.

Fig. 8 Multiple robot engine piston installation system

Fig. 9 Gripper, fixture and other tool design for piston installation


of 47

At the same time as the operations described above for robot II, robot III using the guiding pinsand cap placing gripper picks up the connecting rod cap, moves it under the engine block, andprotrudes the guiding pins through the crankshaft and the cylinder bore of the engine block to receivethe piston connecting rod. Robot II moves the gripper with the piston assembly inserted thereinabove the cylinder bore to be stuffed and then moves the piston assembly into the bore to engage theconnecting rod, leading the tips of the guiding pins into the screw holes on the upper half bearinghouse of the connecting rod. A stabilizing finger is employed to keep the connecting rod in placeduring transportation of the piston assembly to the cylinder bore from the pallet.

Robot II and robot III move cooperatively until the lower surface of the piston skirt is close to theupper surface of the cylinder bore into which the piston subassembly is to be inserted. Then robot IIenables its active searching function to move the subassembly so that the piston skirt finds thatcylinder bore and the piston assembly are inserted into that bore until the lower surface of the gripperjaws touches the upper surface of the cylinder bore. Next, the piston assembly is pushed further intothe cylinder bore by a pusher on the gripper. The third robot with its connecting rod cap placing andrundown device places the connecting rod cap and screws on the inserted piston assembly andfastens the cap onto the connecting rod. The same process is repeated for subsequent cylinder boresin the first row of three cylinder bores for this V-6 engine block.

After the stuffing is finished for the first row of cylinder bores, robot I reorients the engine block sothat the upper surface of the other row of three cylinders bores for the V-6 engine can be stuffed. Thepiston stuffing procedure described above for the first row is repeated to stuff a piston assembly intoeach of the cylinder bores in the second row. Figure 11 is the process flow diagram for the pistonstuffing process with three robots.

In terms of robotic force control searching strategy, it is simpler in piston stuffing than in torqueconverter assembly. There is only one layer here, and spiral search pattern is used. To make sure thesearching does not go too far from the programmed starting position to eventually find the pistonbore (since the chamfer is small on the piston bore, there is no physical limitation for the piston tostay near to the bore during searching), “spring force” is introduced to “pull” the piston and its boretogether during the force-controlled piston stuffing.

SummaryIn summary, robotic force control has been developed and implemented in assembly applications.While started from automotive industry, it has been implemented in general industry. Robotic forcecontrol is an enabler for industrial robots to be used in sophisticated assembly processes. Its

Fig. 10 The unique robotic piston stuffing gripper


of 47

application areas are expanding rapidly. However, since a new hardware device, force/torque sensor,and force control software are introduced, the robotic assembly system with force control is oftenmore complicated and requires the system builder and user to have force control-related knowledge.Furthermore, more robotic assembly process parameters are introduced by force control assemblyprocesses. Assembly parameters such as search pattern, search radius, and insertion force are criticalto the success of the robotic assembly and need to be optimized which will be discussed in section“Assembly Parameter Optimization.”

Assembly Without Force ControlIntroductionSeveral methods have been developed to perform precision assemblies to reduce the design effortsand cost of fixtures that are typically required by high accuracy robotic assembly applications. Thepassive compliance device or Remote Center Compliance (RCC) is an example that allows anassembly robot to compensate for positioning errors due to machine inaccuracy, vibration, ortolerance, thereby lowering contact forces and avoiding part and tool damage. There are many

Fig. 11 The process flow diagram for the three-robot workcell configuration


of 47

research works (Gravel and Newman 2001; Turges and Laowattana 1994; Xu and Paul 1990) doneusing the passive compliance device for assembly or other tasks. However, specific passivecompliance devices have to be designed and manufactured for parts with different geometries,which makes robotic assembly with passive compliance devices more difficult. Also, these devicesare expensive to design and manufacture. To overcome the shortcomings of the passive compliancedevices, robotic assembly using force control (Gottschlich and Kak 1989; Newman et al. 2001) wasthen developed as described in section “Force Control Assembly.” A force/torque sensor is used tomeasure the contact force/torque, and the force/torque signals are used to control the motion of theassembly tool. Since the force/torque control is sensitive to contact with the environment, the toolposition/orientation can be accurately controlled. This makes the force-controlled assembly a goodsolution for robotic assembly. There are many successful stories about robotic force-controlledassembly besides what is described in section “Force Control Assembly,” such as the forward clutchand valve body assemblies, in industrial implementations (Zhang et al. 2004; Gravel and Newman2001; Robotics Application Manual – Force Control for Assembly; Chen et al. 2007). The positionand orientation errors due to the part fixture or location errors can be easily compensated using theforce control strategy. Therefore, the force control method can greatly reduce the requirements of thepart fixtures. Moreover, it can be used for high precision or high accuracy assemblies. However, theforce control method requires extra devices, such as the force/torque sensors and control softwarepackage, etc., which can make the robot control system more complicated and more expensive.Therefore, this section discusses a simpler compliant robot control methodology to perform highprecision industrial assembly that does not require extra equipment or cost yet can be successfullyutilized in many cases.

Position control of industrial robots is very accurate when done with high controller gains.However, contact forces increase rapidly when the robot tooling makes a contact with the environ-ment, making industrial robots difficult to use when limited contact force is needed in high precisionassemblies. By reducing the robot control loop gains, the servos can make the robot compliant to theenvironment, creating a so-called “soft servo” capability. This “soft servo” method can beimplemented to perform some of the high precision assembly tasks where part location errorstypically require the use of force control or RCC methods. The valve body assembly was performedusing both soft servo and force control to demonstrate the performance of both methods.

Fig. 12 The assembly process for high precision assemblies


of 47

High Precision Assembly ProcessA high precision robotic assembly requires a robot to perform assemblies in which the assemblytolerance is better than or close to the robots repeatability. Figure 12 shows a typical method used forhigh precision assemblies. There are three steps in the assembly strategy. Firstly, the robot willsearch locally for the right position of the hole on the valve body to compensate for small errors dueto the vision system accuracy and robot repeatability. Secondly, constant force along the Z directionpushes the accumulator down towards the hole, while other directions are also kept compliant inorder to deal with the orientation errors. Finally, the robot will settle the accumulator down into thebottom of the valve body, and a retrieval force is then applied to retrieve the tool.

A searching method is used to find the exact location of the workpiece. After the part is engagedwith the workpiece, an insertion force is applied to insert the part into the workpiece. Duringinsertion, the tool orientation is changed according to the orientation of the workpiece to avoid jam.Therefore, any developed assembly method needs to be able to locate the workpiece accurately andadjust the tool orientation accordingly.

Soft Servo Control MethodFor a rigid manipulator of n links, the dynamic equation of motion in the joint space is

M qð Þ€qþ C q; _qð Þ _qþ B q; _qð Þ þ G qð Þ ¼ t (2)

wheret 2 Rn vector of applied joint torquesq 2 Rn vector of joint positionsM 2 Rn symmetric positive definite (SPD) manipulator inertia matrixC 2 Rn vector of Coriolis and centrifugal torquesB 2 Rn vector of torques due to friction acting on the manipulator jointsG 2 Rn vector of gravitational torquesWhen there is an external force applied to the robot end effector, the dynamic Eq. 2 becomes

M qð Þ€qþ C q; _qð Þ _qþ B q; _qð Þ þ G qð Þ þ te ¼ t (3)

Wherete 2 Rn is the vector of forces/torques exerted on the environment by the manipulator end effectorexpressed in the joint space.

For model-based control, the model parameters are estimated. Suppose the estimatedcorresponding parameters are M qð Þ, C q; _qð Þ, B q; _qð Þ and G qð Þ , respectively, a typical feedforward PD (proportional plus derivative) controller can be expressed as

t ¼ M qð Þ€qd þ C q; _qð Þ _qd þ B q; _qð Þ þ G qð Þ þ Kpeq þ Kv _eq (4)

Whereep, _ep 2 Rn are vectors of position and velocity errors in the joint space respectively.

Comparing Eqs. 3 and 4, we have

te ¼ Kpeq þ Kv _eq þ Dt (5)

where


of 47

Dt 2 Rn is the vector of forces/torques errors generated by the model estimation error.Assuming Dt is small, the position and velocity errors will balance the external forces/torques

exerted on the robot end effector. Therefore, by decreasing the position control loop gains, the robotposition errors could be increased to make the robot compliant to the environment. Because thecontrol gains of each joint can be tuned individually, a joint soft servo method is formulated. One ofthe advantages of joint soft servo is that the orientation of the tool can be adjusted due to thecompliance of each joint. Therefore, the tool position and orientation can be continuously changedbased on the contact with the environment. The overall control system flow with soft servo is shownin Fig. 13.

There is a switch to select the regular high gain position control and soft servo control. In thenormal position control loop, the regular high gain position control is used. For soft servo control,the search pattern and oscillation patterns are implemented in order to perform high precisionassemblies.

For the X and Y directions, the spiral search method as described in Eq. 2 is used to search the holeposition. For the Z direction, which is the insertion direction, a certain contact force has to bemaintained in the insertion assembly process. To achieve this, a constant error along the Z axis ismaintained:

F ¼ KpzDz (6)

where Kpz is the proportional gain along the Z axis andDz is the set error along the Z axis. These twovalues should be tuned such that the contact force can be maintained in a certain range. Because theinserted part may be jammed during assembly, an oscillation is added to the Z direction to keep thepart from jamming:

Dzs ¼ dz sinwst (7)

where Dzs is the positioning oscillation along the Z axis and dz and ws are the oscillation amplitudeand frequency, respectively.

Experimental ResultsTo demonstrate high precision assembly using soft servo, several experiments were performed. Theassembly process using force control (Details in section “Force Control Assembly”) was alsoimplemented to compare the assembly performance between soft servo and force control.

The valve body assembly was implemented using an ABB IRB140 robot, which is mountedhorizontally on a stand. The software for soft servo and force control was developed on ABB IRC5

Fig. 13 The controller of a robotics system with soft servo


of 47

controller. An ATI Delta force/torque sensor for force control was mounted on the robot end effector,and a suction tool used to pick up the accumulator was mounted under the force/torque sensor. Theexperimental system is shown in Fig. 14.

The valve body was placed in a vise, and, to demonstrate a generic assembly process, theassembly was performed along the horizontal direction.

For the search process using the spiral pattern, the search radius R was set to 1.5 mm and thenumber of turns to 4 (b 2 [0, 8p]). For soft servo, the oscillation amplitude was set to 1 mm and thefrequency to 3 Hz. The settle down point was set to be 20 mm since there is no feedback available.For force control, the search force was set to be 20 N. For the insertion process, the spring forceconstant Kpz was set to 50 N/mm and the settle down force set to 50 N.

A reference configuration was taught by inserting the accumulator into the valve body manuallywhile the force control is active. The position and orientation are recorded as the reference positionand orientation in the robot base frame. Therefore, there are no position and orientation errors if thereference position and orientation are used. However, because there are always fixture errors ina production line, errors were intentionally added to the X and Y axes as disturbance during theexperiments. To compare the performance between the soft servo and force control without bias, theerrors are added along both positive and negative directions along the X and Y axes as shown inTable 1. For the assembly with different errors, the insertion time for both soft servo and forcecontrol is recorded and shown in Table 1.

The data in Table 1 shows that the insertion time using force control is quite close to that using softservo. Therefore, soft servo can be used to perform high precision assembly with cycle timecomparable to that using force control. For larger errors (both the X and Y axes have offsets), theinsertion time is longer. This is because the searching time with big errors is longer. For soft servo,the cycle time is even longer. Thus, the experimental results illustrate that soft servo control is not assensitive as the force control.

Figure 15 shows the valve body assembly process. The accumulator is inserted into the valvebody. Figure 16 shows there are offsets along both the X and Y directions in the tool frame before thesearching method starts.

The contact force signals for both the soft servo and force control were recorded and are shown inFigs. 17 and 18, respectively.

Fig. 14 A high precision robotic assembly system. The suction tool picks up the accumulator and inserts it into the holein the valve body. The robot controller is not shown in the figure


of 47

Fig. 15 The valve body assembly process. The accumulator is inserted in to the valve body

Fig. 16 The valve body assembly process. There are offsets along both the X and Y axes

Table 1 The position offset and insertion time for both soft servo (SS) and force control (FC)

Position offset X Position offset Y Insertion time (SS) Insertion time (FC)

1 0 3.72 3.17

1 0 3.75 3.17

�1 0 3.51 3.15

�1 0 3.53 3.17

0 1 3.80 3.21

0 1 3.82 3.11

0 �1 3.70 3.21

0 �1 3.64 3.20

1 1 4.1 3.20

1 �1 4.72 4.20

�1 �1 5.17 4.67

�1 1 6.30 4.53


of 47

The recorded force signals in Figs. 17 and 18 illustrate that the contact force for both methods arereasonably close. For force control, the reference force is set to 20 N, and the real force signals arequite close the reference force. Because there is no direct force control using soft servo, the contactforce is indirectly controlled by the position offset. The recorded maximum force is about 40 N,which is reasonable for the valve body insertion assembly. Thus, soft servo techniques can be usedfor high precision assembly with reasonable performance compared with the force control method.However, the contact force signals are much more smooth using force control than that using softservo. Also the retract force using force control is almost 0 while there is large retract force whenusing soft servo. This is because soft servo is still based on a position control loop and lower control

Fig. 17 The recorded force signal for force-controlled assembly

Fig. 18 The recorded force signal for soft servo-based assembly


of 47

gains still generate big contact force. The searching time using soft servo is also much longer thanthat using force control, and searching using soft servo is not as smooth as that using force control.

Although soft servo can be used for high precision assemblies, there are some limitations. Theparameters, such as controller gain and maintained contact force during assembly, have to becarefully tuned; otherwise, the parts or the robotic system could be damaged if there was a jamduring the insertion process. The main reason is that soft servo is not sensitive to the contact forcesince it is based on the position errors and not force measurement. Also, the assembly process usingsoft servo is not as smooth as that using force control, and there can be large contact forces generatedduring assembly. The soft servo method is also not as stable as the force control method.While it canbe used to perform assemblies with small errors, if there are large offsets, it will likely fail andgenerate large contact force. Because of these limitations, careful consideration of the intendedsystem is needed before implementing soft servo in an industrial application even though thelaboratory implementation looks promising.

Although there are some significant shortcomings for soft servo assembly methods with compli-ance in all directions, its performance can be quite close to that using force control, especially forsmall initial positioning errors and when using small robots. For large disturbances, a vision systemcould be used to compensate the workpiece location errors, but the added cost and complexity mightbe similar to just using force control.

SummaryIn this section, an assembly method using soft servo is introduced to perform assembly tasks wherepart location errors typically require the use of force control or RCC methods, and the assemblyprocess requires compliance in all directions. The proportional and derivative controller gains arereduced to make the robot compliant when performing the assembly tasks. A search algorithm isdeveloped to find the location of a feature, such as a hole in a part. The tool is kept in contact with thepart while search motion is performed. An oscillation motion is added along the contact directionwhenever the contact friction is severe to prevent binding. Once the tool is engaged with the part,a measured relative mating position (e.g., insertion distance) is used to determine the completion ofthe assembly. Avalve body assembly was used to validate the developed method. Experiments wereconsistently successful when the relative part location errors were within 1 mm, showing that thedeveloped soft servo strategy can perform assembly tasks with small part location errors. Experi-ments with force control were also implemented to compare the performance between soft servo andforce control. It was found that force control methods are much more sensitive to environmentalcontact, and the contact forces can be controlled directly. Conversely, the contact force cannot bedirectly controlled when soft servo is used because it is passive to the contact. Therefore, soft servorequires careful programming and tuning in order to reduce the contact forces; otherwise, damagecould be caused to the products as well as to the robotic system. Also, for bigger part location errors,soft servo either fails to assemble the parts or generates bigger contact forces than force-controlledassembly. Thus, applications of soft servo are more limited, while force control can be successfullyused in most all applications. Further investigation is needed to determine the practical industrial useof soft servo for particular types of precision assembly.

Vision and Force Control-Integrated AssemblyIntroductionMachine vision systems have been widely used to enable a robot to locate the part or subassembly onwhich it is working. In most applications, machine vision systems provide real-time data and livefeedback to guide robots as they go through programmed sequences of operations. However,


of 47

a vision system alone cannot perform high accuracy assembly because the accuracy of the visionsystem is inadequate and the tool may be jammed during assembly. Force/torque control based onthe measured contact force/torque can be added to correct the object position/orientation in responseto the force acting from the environment (Newman et al. 2001). In such a system, an admittancecontrol method (Glosser and Newman 1994) is used to guide the robot tool towards a desireddestination. Such a combination of vision and force control will enable industrial robots to performhigh accuracy assembly in a semi-structured environment.

Some researchers (Morrow et al. 1995) have applied image-based visual servoing techniques toguide the robot to the desired position. However, robots with tooling will block the camera duringassembly in some cases. This prevents the visual servoing techniques from obtaining some real-timeimages, requiring position-based approaches to be implemented in these cases. Position-basedmethods (Jörg et al. 2000) have been studied for assembly too. The exact position of the featureson the assembly parts must be accurately identified. It is well known that accurate calibration ofvision system is very challenging, and disturbances will cause the accurately calibrated systems tohave positioning errors. This makes these methods difficult to use in production environments forhigh accuracy assembly. Yorck et al. (1999) presented a neuro-fuzzy method using visual and forcecontrol. Two camera systems are used for orientation identification. The system is too complicatedand difficult to use in industry.

In this section, an assembly strategy based on combining vision guidance with force control ispresented and demonstrated using a high accuracy assembly process. The vision system does notneed to be carefully calibrated because it is only used for rough positioning. The vision system isused to locate the fixture first. The position of the fixture is sent to the robot controller to guide therobot close to the desired position. Because of calibration errors of the vision system and roboticpositioning errors, it is difficult for the overall positioning errors to be less than the small tolerancenecessary for high accuracy assembly. Therefore, a local searching method based on force/torquecontrol is developed to deal with the errors. Furthermore, the force/torque control algorithm is alsoused to insert the accumulator into the valve body to deal with the position and orientation errorsduring assembly. The methodology developed for high accuracy assembly in a semi-structuredenvironment based on vision, force, and position sensor fusion has been implemented successfullyfor the valve body assembly. The simple vision system calibration also makes the system easy to usein production.

Automated Assembly MethodologyFor high accuracy assembly with fixture errors, the vision system and force/torque control have to beintegrated to perform the assembly task. A vision and force/torque control architecture with

Fig. 19 Control architecture for high accuracy assembly applications with fixture errors


of 47

multilevel accuracy is developed for high accuracy assembly applications with fixture errors.Figure 19 shows the control architecture.

The vision system captures the image of the fixture and computes the X and Yposition errors andthe orientation errors about Z axis. The errors are sent to the robot controller to enable the robot toapproach the desired position and orientation. Because the vision system errors and roboticpositioning error are greater than the assembly part clearance, it is difficult for the robot to achievethe desired position to perform high accuracy assembly. A search algorithm (details in section“Force Control Assembly”) based on the force/torque control is then applied to control the robot toinsert the accumulator into the valve body.

Coarse Control Methodology Based on Vision The coarse control methodology is based on theABB TrueViewTM vision system. Figure 20 shows the TrueViewTM interface.

The TrueViewTM vision system captures the images of a fixture from the camera via a framegrabber. It then computes the position and orientation of the fixture by identifying the predefinedfeatures in the real image. Figure 21 shows the predefined features on a reference image. Thesefeatures are defined by the user. One feature is enough to define the position and orientation of thepart; however, the identification accuracy can be increased by defining more features. A coordinatesystem in the image frame is also defined as shown in Fig. 21.

To control the motion of the robot, the coordinates in the image frame have to be transferred intothe robot frame. This is done by the ABB TrueViewTM automated calibration feature. This tool usesa pattern, either in dot format or square format, to calibrate the system. Figure 22 shows the dotcalibration patterns.

The image of the calibration pattern is captured and analyzed automatically. Figure 23 shows thecalibration interface of the ABB TrueViewTM system. In the image frame, a 7 � 7 dot pattern ischosen. In the robot world frame, the positions of the three dots of the origin Po, the X axis PX, andthe Y axis PY (as shown in Figs. 22 and 23) are measured by teaching the robot.

The coordinates of the origin, PX and PY of the defined coordinate system, are input into theinterface. The calibration is then performed automatically, and the calibration matrix is identified.

Fig. 20 The TrueViewTM vision system interface


of 47

During assembly, the image of the fixture is captured. It is then analyzed and the predefinedfeatures are identified in the image. From the position and orientation of the predefined features, theXY position and orientation about Z axis of the valve body are calculated. Using the calibrationmatrix, the position and orientation of the valve body in the robot world frame are computed andused to control the robot to approach the desired assembly configuration.

Fig. 22 The dot calibration patterns used to calibrate the relationship between the image frame and the robot worldframe

Fig. 21 The defined features and the coordinate system


of 47

Fine Control Strategy Based on Force/Torque Control Because of the camera resolution,distortion, and calibration errors, a robot guided by the vision system will have some smallpositioning errors. For high accuracy assembly, these small errors will cause the assembly to failbecause traditional industrial robots are very stiff. Therefore, a fine control strategy based on force/torque control is developed for the assembly as described in section “High Precision AssemblyProcess.”

The robot holding the accumulator starts to search for the right position of the hole from the centerof the spiral pattern while the force along the Z axis is kept at a constant value.When the accumulatoralong the Z axis changes by a set threshold, it is inserted into the hole on the valve body, and thesearch is terminated. If the robot reaches the boundary without finding the desired position, it willsearch back to the center of the spiral pattern. This search process is repeated several times until theaccumulator is inserted into the hole. The initial position error of the tool should not be too big. If theposition error is too big, the search process will not find the desired position, and the search will beterminated automatically after a few trials. The number of trials and turns in the spiral pattern can bedefined by the users.

For the insertion process, the approximate insertion distance is known. Therefore, based on theinsertion distance, the reference force along Z axis can be calculated and used to control the motionof the robot. One of the advantages of this control strategy is that there is no large initial impactcontact force and the cycle is minimal. The tool orientation is adjusted based on the measured torquesuch that the accumulator can be inserted without failure.

Once the tool approaches the settle down point, a constant force is set to make sure theaccumulator is properly inserted. After the accumulator reaches the bottom, the contact forcealong the Z axis will increase.When the contact force reaches the set value, the insertion is complete,and the tool is retrieved. The assembly is then complete.

Implementation of High Accuracy Assembly with Fixture ErrorsThe high accuracy assembly is implemented using an ABB IRB140 robot. An ATI Delta force/torque sensor is mounted on the robot end effector. A suction tool is used to pick up the piston and is

Fig. 23 The calibration interface of the ABB TrueViewTM system


of 47

installed under the force/torque sensor. A Pulnix TM-200 camera is mounted on a frame that isstationary to the robot base frame. The experimental system is shown in Fig. 24.

The valve body is placed on a table. The distance from the camera to the table surface is about550 mm. The field of view is about 180 � 135 mm. The resolution of the camera is 640 � 480.Therefore, the accuracy of the camera system is about 0.33 mm. This accuracy is inadequate for highaccuracy assembly. The ABB TrueViewTM system is installed on a PC with a Meteor II framegrabber. The PC and IRC5 robot controllers communicate via Ethernet. The fixture position errorsare computed by the ABB TrueViewTM system and sent to the robot controller. For the searchprocess using the spiral pattern, the search radius is set to be 1.2 mm, and the number of turns is4. The search force is set to be 20 N. For the insertion process, the spring force constant is set to be50 N/mm. The settle down force is set to be 50 N.

In order to improve the accuracy of the assembly system, the position errors and orientation errorsrelative to a reference robot configuration are used to calculate the robot position and orientation.A reference configuration is taught by inserting the piston into the valve body manually while theforce control is active. The position and orientation are recorded as the reference position andorientation in the robot base frame. After that, the robot moves to the “Home” position withoutblocking the image of the fixture. The TrueViewTM system will capture the image, analyze theimage, and output the part coordinates in the robot frame using the calibration matrix. These XYcoordinates obtained from the image system may be different from the taught reference position andorientation in the base frame. The two reference position and orientation differences are calculated.When the valve body moves to different places, the differences are used to compute the robotposition and orientation in the robot base frame. This method can decrease the calibration error.

Two sets of experiments are performed to verify the new vision and force control strategy. In thefirst set of experiments, the valve body is fixed on the table using a clamp. Figure 25 shows theassembly when the fixture is set at different places, and the assemblies are successful. A series ofexperiments were performed without failure. In the second set of experiments, the valve body is freeto move on the table to simulate the assembly without fixture. Figure 26 shows the assembly atdifferent places without fixture, and the assemblies are successful. A series of experiments wereperformed without failure. Therefore, the developed strategy can be used for the high accuracyassembly in manufacturing.

Without a combined vision and force control strategy, the high accuracy assembly is hard to beimplemented. Figure 27 shows the assembly without vision guidance. The position errors are toolarge, and the robot cannot insert the accumulator into the valve body.

Fig. 24 The high accuracy assembly system. The suction tool will pick up the piston and insert it into the hole on thepiston body


of 47

Figure 28 shows the assembly without force/torque control. The accumulator is jammed at the topof the valve body and cannot be inserted into the valve body.

Fig. 26 The high accuracy assembly without fixture

Fig. 25 The high accuracy assembly with fixture


of 47

SummaryIn this section, a vision and force control-integrated robotic assembly technique is discussed. A highaccuracy assembly system is successfully implemented based on a combined vision and force/torquecontrol strategy in a semi-structured environment. The vision system guides the robot tool near thedesired position. A force/torque control strategy is then applied to search for the desired position andperform the assembly. The system can also be applied to other similar assembly processes such asmanual clutch assembly, torque converter assembly, etc.

Assembly on a Moving Production LineIntroductionRobotized moving line assembly system (assembly is performed while the part is moving) is anintelligent robotic system based on vision, force, and position sensor fusion while the object ismoving randomly on a platform. The robot can track the moving object while performing theassembly process. Compared to the stop station assembly (assembly is performed when the part isstationary), the moving line assembly can save huge amount of time and resource. However,integration of visual servoing and force control in industrial robotic systems is still a very challeng-ing research area, especially when the system is installed for trim-and-final assembly processes inautomotive manufacturing because a small damage to the parts could cause a big loss.

In patent 6,886,231,B2, a stop-assembly-station system is described. The vision system is used toidentify the position/orientation of the part. The assembly is carried out by using the vision system

Fig. 28 The high accuracy assembly failed without force/torque control

Fig. 27 The high accuracy assembly failed without vision guidance


of 47

alone. A vision and force control -based stop station assembly system is described in patent6,141,863. The force control is used to execute the insertion task, and the vision system is used toidentify the position/orientation of a part. In patent 5,040,056, a part picking up system is presented.The part position/orientation is identified by the vision system, and the data are transferred to therobot to pick up the part. Although the conveyor is moving, the vision system is used only forposition/orientation identification. Therefore, no visual servoing involved. There are many researchpapers about moving line tracking, such as Cho (2005) and Shin (2006). These systems use onlyvisual servoing to track the motion of the parts to perform the assembly tasks. In the actualapplications, especially in the trim-and-final assembly, the malfunction of the visual servoingcould damage the final products, which is very costly. Beaten et al. (Beaten and Schutter 2002;Beaten et al. 2002) developed a hybrid vision and force control strategy to follow the contour ofplanar surface. Xiao et al. (2000), Chang et al. (2002), and Olsson et al. (2002) presented methods tofollow a curve on a surface while maintaining a certain contact force. Although these methods candeal with the feature following based on hybrid control strategy, the applications in the movingobject tracking while performing certain tasks are not discussed. Nelson et al. (1995) discusseddifferent methods based on vision and force control, such as hybrid control, traded control, andshared control methods. Interesting experimental results are presented regarding the performance ofthese methods. However, the force control is only used to maintain a certain contact force. Althoughcombination of vision and force control strategy has been widely studied, no practical implemen-tation to perform complicated assembly processes has been done to perform assembly using theintelligent robotics system based on the synergistic combination of vision, force, and position as thefeedback information.

In actual applications, the part on the moving assembly line may move randomly along the X, Y,and Z axes. 3D vision (stereo vision) system can be used to track the motion of the part (Yoonet al. 2006). The processing of stereo information enables a robot to determine the position andorientation of an object in the robot coordinate frame. However, an accurate calibration between thecamera coordinate system and robot coordinate system has to be realized. Because the stereo visionsystem requires high-quality camera, accurate calibrations, and high computation power, theexisting systems are costly, error prone, and not robust enough for daily use at the workshop.Also the complicated computation for determining the position/orientation of an object makes thesystem difficult to be implemented in real-time applications. Furthermore, the visual servoing alonecould cause damages to the final products if the vision system malfunctions. Therefore, the roboticsystem must be compliant to the environment during the assembly process. Hence, force controlcombined with a simple vision system is a better choice for industrial applications. To perform anassembly using a single camera system, the X and Y axes errors of the moving part can becompensated by visual servoing; however, the Z axis errors are difficult to compensate using onlyone camera. Thus, force control is applied to control the motion of a robotic system to perform anassembly task along the Z axis.

This section discusses an assembly technology while the object is moving using visual servoingand force control. The wheel loading process, which assemblies the wheel onto the wheel hub ofa vehicle on a moving assembly line, is used as an example to demonstrate the developedtechnology. A working system has been set up, and experiments of loading wheels onto the wheelhub were implemented and performed successfully. The experimental results demonstrate that thedeveloped technology can be used for assembly while the part on the assembly line is movingrandomly.


of 47

Moving Line Assembly MethodSystem Structure The developed moving line assembly system can track the moving object whileperforming the assembly process based on the vision, force, and position sensor fusion. Figure 29shows a moving line assembly system. The object (part) is moving while the conveyor is moving.The motion of the conveyor is typically random with the velocity change along both the X andY directions as shown in Fig. 29. Hence, the robot has to track the motion of the moving part alongboth the X and Y directions in order to perform the assembly processes. This is different from thetypical conveyor tracking process which only the motion along the conveyor moving direction iscontrolled.

Since the system involves the vision system, force control system, and the robot control system,the system structure is discussed first.

The control structure is shown in Fig. 30. The vision system processes the captured images andsends the position signals to the robot controller (IRC5, an ABB controller) to control the motion ofthe robot to follow the motion of the object. The force sensor measures the contact force to adjust the

Fig. 30 The control system structure. IRC5 is an ABB robot controller

Fig. 29 A moving line assembly system tracks the motion of the object while some assembly processes are performed


of 47

motion of the robot. The force control also makes the robot compliant to the environment to avoidthe damage to the contact object.

In the vision system, a predefined feature on the moving part is identified by a camera. Theposition error of the feature compared with the desired feature in the camera frame is computed tocontrol the motion of the robot as shown in Fig. 31. Since only one camera is used, the robot can onlytrack the motion of the moving part along the X and Y directions. To perform an assembly, the robothas to be controlled to approach the moving part. Therefore, force control along the Z axis is appliedto control the motion of the robot to perform an assembly process. Since the visual servoing alonecould cause damages to the final products if the vision system malfunctions, the force control alongboth the X and Y axes is also enabled to make the robotic system compliant to the contactenvironment.

Several coordinate frames are involved in the system: the conveyor frame (work object frame), thefeature frame, the camera frame, the tool frame (gripper), the force sensor frame, the robot baseframe, and the world frame. Most of the frames are common except the feature frame. The featureframe is defined on the center of the feature and is parallel to the work object frame. Later on, theseframes are used to describe the algorithms.

Control System The motion of the robot is controlled by the robot controller based on the sensorinputs. The control system structure is shown in Fig. 32.

Fig. 31 The actual identified feature and the desired feature in the camera frame

Fig. 32 The control system structure of the moving line assembly system


of 47

After the feature error is identified, it is input to the visual servoing controller to correct the featureerror. The force sensor measures the contact force and torque. The desired force and torque arecontrolled to be zero except the Z axis that is controlled to move the robot towards the moving part.The force and torque errors are sent to the force controller to control the contact force. This is a trade-off between the visual servoing and force control that is achieved by the controller gains. Thesecomputed velocity errors are input to the robot controller to regulate the robot velocity to achieve thedesired performance.

The reference velocity vref is generated using the actual gripper (tool) velocity v, the velocityerrorDvI computed from the vision system in the tool frame, and the velocity errorDvf from the forcecontrol system in the tool frame, as described in the following equation:

vref ¼ vþ DvI þ Dvf

where

DvI ¼ KI X d � Xð ÞDvf ¼ Kf Fd � Fð Þ

Fd and F are the desired force and actual measured contact force, respectively, in the tool frame.KI

is the visual servoing coefficient and Kf the force control coefficient. Xd and X are the desired featureposition and actual feature position, respectively, in the camera frame. Because the velocity errorfrom the vision system is computed in the tool frame, the transformation matrix from the cameraframe (the image frame) to the tool frame has to be identified. Since the camera is moving towardsthe moving object, the scale of the feature is keeping changing related to the distance from thefeature to the camera. Therefore, the desired feature position is updated on the real time.

Experimental Implementation and ResultsTo validate the developed method, experiments were implemented to perform an assembly processwhile the object (part) is moving on a conveyor. AnABB IRB6400 robot with force control was usedas shown in Fig. 33. The wheel loading, a representative assembly process, is used to validate thedeveloped methodology.

Fig. 33 The wheel loading system. The wheel is assembled on the wheel hub while the wheel hub is moving ona conveyor


of 47

Figure 33 shows the wheel loading system. The robot is controlled to automatically assemble thewheel into a wheel hub based on the synergistic combination of vision, force, and position as thefeedback information to control the robot motion. The steps are as follows:

1. The 2D vision system identifies the position/orientation of a wheel and picks it up as shown inFig. 34.

2. The robot moves to the first position (taught manually). The camera will take an image when themoving assembly line (conveyor) triggers the first trigger. After image processing, the orientationpattern of the bolts on the hub is identified. The robot additional axis will rotate to match the holeson the wheel to the bolts on the hub.

3. While the additional axis is rotating, the robot moves to the second position. When the secondtrigger is triggered, the vision system will identify the speed of the moving object and track it. Atthe same time, force control is enabled along all directions. The robot will move towards the hubbecause of force control along the Z axis.

4. Once the contact force along the Z axis reaches a certain value, the assembly completes and theforce control along the negative Z axis will retract the robot.

5. The force control along all directions will keep the system safe under abnormal conditions. Forexample, when the vision system sends wrong signals, the force control will balance the wrongsignals and will not cause damage to the system and the product.

There are several calibrations involved in the wheel loading process. First, the holes on the wheeland the studs on the hub have to be matched. This is done by teaching. The tool location is then

Fig. 34 The robot picks up a randomly located wheel. (a) The wheel is randomly located on a feeding table. (b) Thevision system automatically identifies the location of the wheel to control the robot to pick up the wheel


of 47

recorded, and the transformation between the tool frame, the feature frame, and the camera frame isthen computed. The wheel is then moved and dropped on the wheel feeding table. The robot ismoved to a position to capture the image of the wheel, and the reference position and orientation arecomputed. The camera is then moved to several positions to calibrate the transformation matrixbetween the wheel object frame and the tool frame. These data are used to compute the wheellocation when the wheel is loaded on the table randomly.

Figures 35 and 36 show the assembly process. The robot tool grips the wheel and tracks themotion of the wheel hub. The wheel is then installed on the wheel hub using force control. The forcecontrol is enabled along all direction to make the system flexible to the environment.

Figure 37 shows the velocity along the X and Y directions recorded from the robot controller. Thevelocity is changing randomly along both the X and Y directions. The vision system is tracking thefeature and input the velocity errors to the robot controller to control the motion of the tool.

Fig. 35 The robot tracks the motion of the wheel hub while the force control enables the tool approaches the wheel hub

Fig. 36 The wheel is successfully installed on the wheel hub. The assembly process is completed


of 47

SummaryThis section presents an intelligent industrial robotic system that can perform assembly tasks withhigh autonomy and adaptability to the environments. The assembly is performed while the object ismoving based on the visual servoing and force control strategy. The wheel loading process, whichassemblies the wheel into the wheel hub of a vehicle on a moving assembly line, is used as anexample to demonstrate the developed technology. Experiments were performed successfully, andthe results demonstrated that the developed technology can be used for assembly while the assemblyline is moving randomly. Since huge amount of time and resource can be saved using the developedintelligent robotic system, this innovative technology will have great impact in the automotiveindustry, especially when the labor cost is becoming higher and higher.

Fig. 37 The recorded tool velocity. The object is moving along both the X and Y axes


of 47

Assembly Parameter Optimization

Off-Line Parameter OptimizationIntroductionWith the high demand of flexible manufacturing automation, industrial robot applications have beenexpanded into processes with contact force such as assembly andmachining. In conventional roboticassembly applications, a passive compliant tool is commonly used to compensate the robot positionerrors. It works for assembly processes with relatively simple and loose insertions. For processeswith high precision and tight tolerance requirements, robot force control has been developed andimplemented. This recent advance in industrial robot control gives a “touch” sensing of theindustrial robots and permits an entire new class of robot behaviors and applications. The newrobot behaviors are possible due to the incorporation of force sensor into the robot control system.The force control robot technology enables robotic automation applications that mate parts togethersuch as gear meshing, spline insertion, clutch hub assembly, and surface grinding following complexcurved geometry. Gravel et. al. (2001) give an overview on the research and development ofassembly applications using robots with force control. Robotic assembly application projects thatinvolved universities and robotic manufacturers were described. Robotic assembly witha complicated and heavy (up to 75 kg) transmission component, torque converter, was successfullyperformed (Zhang et al. 2004). The system has been installed and run in multiple production lines forseveral years. Besides torque converter assembly, transmission valve body assembly and enginepiston installation have also been investigated. The papers (Wang et al. 2008; Zhang et al. 2007)dealt with force control technology and applications. More details in force control assemblytheoretical development and application are described in section “Force Control Assembly.”

On the other hand, robot force control introduced complexity and uncertainty to the robotprogramming, control parameter setting up, andmanufacturing process quality. The force-controlledrobot behaves differently for different contact force conditions resulting from the manufacturingvariations of the assembled parts, fixture, and environment disturbances on the manufacturing floor.One of the most recognizable behavior differences from position-controlled robot is that the robotmotion cycle time is no longer a predetermined value in force control. Normally, it will be distributedin a statistics manner for robotic assembly processes. And the mean and standard deviation of theassembly cycle time and FTT rate become the measurement and the optimization objectives ofa force-controlled robot assembly system. In order to optimize the performance of robotic assembly,the process needs to be first parameterized.

Since the introduction of force control, in which actual robot path depends on not only theprogrammed position but also the interaction force between the assembled parts/components, theoptimization has become more difficult. Although the search-based assembly strategy offers simpleand robust solution that is favored in the industrial environment, it is still difficult to find the optimalparameter settings for the search motion. An optimization procedure is therefore much needed forthe search-based strategy. On the other hand, the search-based strategy is very well parameterized byits search motion parameters and therefore is very suitable for optimization. The optimal set ofrobotic (force control) parameters is often obtained either by trial and error or off-line analysis tools.It is tedious and time consuming. Because of the statistical nature of the assembly task andincreasing popularity of design of experiment (DOE) (Montgomery 2005) in manufacturing qualitycontrol, an off-line DOE has been used in the robot assembly parameter optimization to assist thesystem setup (Robotics Application Manual – Force Control for Assembly). The paper (Gravel2007) introduced a systematic method to obtain a set of optimal assembly parameters by DOE. Afterthe assembly parameters to be optimized are selected, the experiment is designed and coded into


of 47

robot program. The program is executed on a robot controller, and the result data is stored in a datafile. The data is processed by a separate computer with a statistic analysis software package such asMinitab. This whole parameter optimization process often needs to involve a DOE expert, a robotprogrammer, and a robot operator so it is difficult to be used in manufacturing floor in massiverobotic assembly production environment.

Theoretical DevelopmentParameterizing the Robotic Assembly Process In order to optimize the robotic assembly param-eters, the process needs to be first parameterized. Based on the nature of common assembly types,assembly processes are categorized into cylindrical, radial, and multistage insertion/assembly in thisinvestigation. Cylindrical insertion includes assembly applications with a cylinder and a smoothbore, such as transmission torque converter assembly shown in Fig. 4 and piston stuffing in Fig. 6.

Spiral search is used as primary pattern. Radial insertion includes assembly applications witha “toothed,” “splined,” or “gear”mesh, such as a forward clutch assembly and spline gear assembly.Rotational search is used as primary pattern. And multistage insertion which requires the combina-tion of cylindrical and radial at different stages includes assembly applications such as a torqueconverter assembly, half shaft-differential assembly, and the like. The combination of rotational andcircular motion is often used as search pattern. Besides, hopping, oscillating along the insertiondirection, can help a lot to overcome the contact friction force and the sliding between the loosingparts. The cylindrical, radial, and multistage insertion types cover most of the real-world assemblyapplications. For each insertion/assembly type, certain search pattern is applied to perform theassembly process. Correspondingly, there are certain robot force control parameters that are relatedto each search pattern. As an example, the robotic assembly parameters for multistage type used intorque converter assembly are described in this Section. A RAPID system module, AsmWareBase,has been written to carry out the assembly process parameterization. The AsmWareBase module canbe called in samemanner as standard RAPID instructions. There are two sets of assembly parametersin the specification. One set of the parameters is called Set Parameter. They are related to the searchpattern transition, process action, and insertion termination and normally setup at the beginning ofthe robotic assembly engineering process, as listed below:

• Insertion Distance – distance between the insertion starting position and the end position• Engage Distance – distance between the insertion starting position to the position where the parts

are engaged• Time Limit – time limitation before a search process is terminated or re-performed;• Max Num Try – maximum number of trials before an unsuccessful insertion is claimed• Use Timeout Act – a flag that signals whether or not to use a customer-specified timeout handling

method• TO Action Num – timeout handling routine number• Use Force Cond – a flag that signals whether or not to use force control insertion termination• Cond Force Value – condition force value used in force control insertion termination• Use IO Action – a flag that signals whether or not IO action is used at end of the insertion• IO Action Num – the IO number used in the action above• Use Force Retreat – a flag that signals whether or not the force control retreat is used• Retreat Force – the force value used in the force control retreat


of 47

Another set of the parameters, which are related to the performance tuning of the robotic assemblyprocess, is called Tune Parameters. Tune parameters are insertion/assembly type specific. For thetorque converter assembly, the tune parameters may be:

• Search Force – max value of the searching/insertion force used in the tool Z direction• Rotation Speed – the rotation speed used in the insertion around the tool Z axis• Rotation Angle – the maximum angle value used in the above rotation motion• Circular Speed – the circular searching speed in the XY plane of the tool frame• Circular Radius – the circular radius used in the insertion corresponding to the circular speed

above• Force Amp – the amplitude of the hopping force in the Z direction of tool frame• Force Period – the period of the hopping (oscillation) force

The combination of the given assembly parameter and contact force defines the actual assemblypath. An example of the radial-type insertion reference path is shown in Fig. 38.

Fig. 39 DOE process analysis diagram

Fig. 38 The radial-type insertion reference path


of 47

Design of Experiment Methodology DOE is a powerful technique used for exploring newprocesses, gaining knowledge of the existing processes, and optimizing these processes for achiev-ing optimal performance. DOE’s design and analysis theory is used in this investigation. The book(Montgomery 2005) gives the design and formulation detail.

As illustrated in Fig. 39 for a particular process, the controllable variables such as productionbatch are defined, the uncontrollable variables are ignored, the inputs are varied in a designedmanner, and the output characteristics are measured.

A fractional factorial test is used to find the most influential parameters first, and then full factorialtests are used to result in a set of optimized parameters. In this robotic assembly parameteroptimization, for example, the input parameters are Search Force, Rotation Speed, RotationAngle, Circular Speed, Circular Angle, Force Amp, and Force Period. The output variables areAssembly Time and FTT rate.

The DOE Analysis of Variance (ANOVA) method, as shown in the diagram in Fig. 40 is used inthe experimental data analysis.

Suppose there are a levels of a single factor that are to be compared, the observed response fromeach of the a levels is a random variable. The data would be yij, representing the jth observation takenunder factor level i. There will be, in general, n observations. Then

yij ¼ mþ ti þ eiji ¼ 1, 2, . . . , aj ¼ 1, 2, . . . , n

�(8)

wherem is a parameter common to all levels called the overall mean.ti is a parameter unique to the ith level called ith-level effect.eij is a random error component that incorporates all other source of variability in the experiment.

Equation (3.1.1) is called the single-factor analysis of variance. If the experiment design isa completely randomized design, the objective is to test appreciate hypotheses about the treatmentmeans and to estimate them. For hypotheses testing, the model errors are assumed to be normallyand independently distributed random variables with mean and zero and variance s2. The variances2 is assumed to be constant for all levels of the factor. This implies the observations:

yijeN mþ ti, s2� �

(9)

and that the observations are mutually independent. DOE is selected as the optimization method due

Fig. 40 DOE analysis of variance diagram


of 47

to its popular acceptance by the industry, although other nonlinear optimization methods such asgenetic algorithm and simulated annealing can be also applied. Due to the randomness of theassembly strategy and the part positional uncertainties, it is necessary to run at least 10 replicatesof each experimental trial in order to get a reliable performance measure. The performance measurehere is chosen as a statistic of the cycle time:

P ¼ mþ ws � sþ wr � m � r (10)

wherem is the mean cycle time.s is the standard deviation of the cycle time.r is the failure rate after the repeated trials.ws and wr are the weight of the standard deviation and the failure rate.The motivation behind this performance measure is that a production process should not only

have lowmean cycle time and failure rate but also have low variability in the process. In addition, forsafety reasons all the parameter values should be properly bounded.

The DOE process can be performed at each substage, as illustrated in Fig. 41.The optimization of each substage starts with the manual selection of the optimizable parameters

and their bounds. The purpose of the initial screening is to reduce the number of parameters so thatthe final optimization step can be performed more efficiently. The typical DOE-based parameteroptimization steps normally are as follows:

1. Screening: using fractional factorial experimental design matrix to generate the experimentalruns, which consists of a carefully chose fraction (subset) of the experimental runs of a fullfactorial design based on the sparsity-of-effects principle. Then the collected data are subjected tostatistical analysis to identify the most influential parameters. Figure 42 shows a typical Pareto

Fig. 41 DOE optimization procedure


of 47

plot from screening process. The higher the value of the parameter in the plot, the more influentialthe parameter contributes to the output performance.

2. Optimization: based on the screening result, using full factorial experimental design matrix togenerate the experimental runs, which consists of 2-level or 3-level discrete possible values ofmost influential parameters. The results of experimental runs are analyzed to find the optimalparameter set.

3. Verifying: verifying the optimized parameter set by running a number of experiments andchecking on the distribution of the objective metrics. The number of parameters to be optimizedand the level of values for each parameter can be varied widely based on the sensitivity of theoptimization goals to the parameter change, the available number of tested parts, and the cost ofthe experiment.

ExperimentWith the development of the on-pendant assembly parameter optimization tool, tests have been doneon both transmission valve body and torque converter assembly applications. This section presentsthe testing result using a torque converter assembly. As shown in Fig. 43, a Ford 6R60/80transmission torque converter is used in the test.

Fig. 43 6R60/80 transmission torque converter assembly process

Fig. 42 The Pareto plot from screening process


of 47

After basic robot force control setup, the initial assembly parameter values are chosen byanalyzing the torque converter drawings, measuring part dimensions directly, and/or observingmanual assembly process. For this particular case, the assembly process is divided into two stages.The first stage is from the start position all way to matching spline gears until the pump gear. In thisstage rotational search range would be relatively small (about 45�), and circular search with certainradius is necessary to deal with the starting XY position errors. Default offset (10 %) is added andsubtracted from the nominal values to get the lower and upper boundaries. Figure 44 shows theinitial parameter values and lower and upper boundaries for the first stage.

When initial parameter values are set and tested, screening run can be performed. PB-8 or PB-12(Plackett-Burman method) could be used for the seven-parameter screening. The same experimentcan be repeated ten times for better accuracy in statistics sense. After the screening run is completed,experimental data are processed by pressing the “Update” button, and the screening result isdisplayed in the order of the most to the least influential parameters to the assembly cycle time.After the screening is done, the first three most influential parameters will be automatically selectedfor the optimization process. Figure 45 shows the setup for parameter optimization.

If a parameter other than the default one is to be optimized, the box on the far right side can bechecked. Only can three parameters be optimized at a time. Figure 46 shows a typical optimizationresult.

Fig. 44 The initial parameter values and lower and upper boundaries for the first stage

Fig. 45 The setup for parameter optimization


of 47

The first three most optimal parameter sets are displayed along with their success rate if multipletests have been performed for the same experiment. Main effect plot can be seen by clicking thecorresponding icon. Figure 47 shows a typical main effect plot. From the main effect plot, the trendof the optimal parameter value of a particular parameter can be figured out.

To verify the mean and standard deviation of the assembly cycle time and success rate ina relatively large member of samples, a verification experiment is conducted. Up to 100 tests canbe run, and resulting data can be processed by pressing Calculate button on the verification resultpage shown in Fig. 48.

Figure 49 shows a typical assembly cycle distribution plot for a torque converter assembly. Theassembly cycle time mean and the standard deviation are 7.56 and 2.65 s, respectively, in thisexample.

If the verification result is satisfactory, the parameter set can be used in production. Otherwise theparameters can be used as new initial parameters for the next round of parameter optimization untilthe desired parameter values are obtained.

SummaryThe assembly parameter optimization tool has been used in finding a starting parameter set for a real-world robotic assembly cell. It has been proven that using the optimization tool, the robotic assemblycell programming time including obtaining the optimal parameter set can be reduced from weeks todays. Realizing that the optimal parameter set is not only related to the production batches but also

Fig. 46 Optimization result

Fig. 47 Main effect plot


of 47

related to the assembly conveyor and pallet accuracy, the robot gripper design, and manufacturingtolerance as well as the assembly starting position, the optimization result is robot cell andinstallation specific. In the applications which use same type of parts, same robot models, andsame assembly programming strategy, the optimal parameter values could be close but often notexactly the same. So the optimization ought to be done on the production unit and on themanufacturing lines. To expand the parameter optimization concept, the next step is to optimizethe assembly parameters automatically online (in production) which will be described in the nextSection.

Online Parameter OptimizationIntroductionDOE is faster than exhaustive parameter searches and is driven towards a systemic optimal solution(Siciliano and Villani 1999) but usually requires several professionals to construct and implement.At the early stage of the optimization work, in order to find the optimized parameter for setting up thesystem to meet the production, the experiments were designed off-line using Minitab on a PC andprogrammed into robot motion program. The robot program runs on the actual or close to actualproduction environment and the result data are collected. The data file is then imported into Minitabfor analysis. This process is often needed to be done for several iterations. There are severalprofessionals needed to be collaborated in performing this optimization task, includinga manufacturing quality control expert to analyze the assembly process and design proper DOE

Fig. 48 Verification result

Fig. 49 Assembly cycle distribution plot


of 47

experiments, a robot programmer to code the design into robot program, an operator to execute theprogram, and then the quality control person to utilize the data to design a new experiment. To assistthe optimization process, improve the efficiency, and make the optimization process down to themanufacturing floor, a robotic assembly optimization tool, running in an auxiliary processor in theteach pendant of the robot, is developed (Xu and Paul 1990). However, in the production of roboticassembly, for example, torque converter and transmission assembly, even with the optimizedparameters, the performance of the robotic assembly system is deteriorated when manufacturingenvironment changes such as the variation of geometrical dimension of a part and tool (the locationof feature on part, the size of the feature, the dimension of the tool, etc.); the changes of position andorientation of a part, fixture, or robot; and the changes of properties of a part (weight, springconstant, etc.). The proposed method applies DOE-based parameter optimization in production toautomatically adapt to the manufacturing environmental changes in robotic force control assemblyand optimize the productivity of the assembly. The challenge in parameter optimization in produc-tion is to minimize the disturbance of experimental trials and balance the cost of the number ofexperiments and optimization goals. Because the robotic torque converter assembly cell operates ina continuously running production line, a certain throughput is designed and maintained in order tohave smooth production. There are only limited buffers in production line to accommodate the smallvariation of the assembly cycle time. On the other hand, in order to find the optimized parametersand adapt to the manufacturing environment changes, the values of the current production param-eters have to be varied to perform the DOE experimental trials. Correspondingly the performance ofthe robotic assembly is varied. So the keys to the success of automatic in production parameteroptimization are the following:

1. Limit the interference of the optimization process to production. The proper algorithm should bebuilt into the optimization process to control how to vary the parameters, when to switch intooptimization, and when to switch back to production.

2. Also the access to the result of optimization experimental trials, the analysis of the result, and theupdate of the production parameters should not disrupt the production.

3. Moreover, the cost of the experiment and optimization goals needs to be balanced. The number ofthe experimental trials and the number of the optimization iterations need to be limited. Althoughin theory the more trials the less sensitive the result is to the random noise in the process and themore iterations of optimization the close the result is to the optimal solution, the variation of theassembly cycle time during the optimization process will make the overall performance of theproduction worse. More important the goal for the online parameter optimization is NOT to findthe best set of parameters from the thousands of trials happened long time ago; the goal is toquickly catch the changes in the parts and fixtures and optimize the system to adapt to thesechanges. The manufacturing environment probably changes again before the optimization iscomplete.

Theoretical DevelopmentFigure 50 illustrates the flow chart of the online DOE-based parameter optimization process, whichcontains three major functional tasks: (1) process monitoring, (2) parameter optimization, and(3) parameter verification.

The online DOE-based parameter optimization removes the screening process to simplify thestandard DOE process. The assumption is that the screening has been done in system setup process,and the rank of the influence of the parameters does not change greatly. In the normal production,


of 47

which is a stable process, this is a valid assumption since the manufacturing environment will onlychange gradually. If the manufacturing environment changes dramatically, the production becomesunstable. The DOE-based parameter optimization should not be used to recover the productionperformance. The intervention from the professionals is expected to investigate the root cause of thechanges.

An algorithm is built into the optimization process to compare the verification result with thebaseline result. The baseline result is the performance of production parameters and the performanceof the optimized parameters in the previous optimization iteration. If the improvement is notsignificant, as defined by the user in terms of % improvement in the cycle time metrics, theoptimization process is complete.

ExperimentBased on a real-world transmission torque converter assembly production process, investigation andanalysis are performed in production. An on-pendant robotic assembly parameter optimization toolis developed to implement the online DOE-based parameter optimization technique. The optimiza-tion tool, running in an auxiliary processor on the teach pendant of the robot, applies full factorialexperiments on the most influential parameters. Then the results are subjected to statistical analysisto find the optimal parameter set. The optimized parameter set is verified through running a numberof experiments and checking on performance of the force control assembly to adapt to the changes.The software structure is illustrated in Fig. 51.

In this software package, the robotic assembly process (using torque converter assembly asa development platform, but the tool can be used for various assembly processes of other types)has been firstly parameterized into operator-understandable terms and parameters such as startingpoint, assembly stage, insertion distance, timeout limit, max number of trials, searching force,rotation speed, rotation angle, force amplitude, force (sin wave) period, and so on. A robot program(ABB robot programming language, RAPID) is written to perform a particular assembly production.Another RAPID program module is to convert the process-related parameters into robot forcecontrol parameters, generate the DOE experiment design matrix to vary the value of the optimizingparameters, and collect the experimental trial results. Then, a DOE design and analysis function hasbeen simplified specifically for the robotic assembly applications and coded into a C# library torealize the DOE design and analysis on the touch-screen robot teach pendant (so-called ABBFlexPendant).

Both RAPID production program module and DOE module are running in the ABB IRC5 robotcontroller. The DOE RAPID module seamlessly interacts with the force control assembly produc-tion program to switch the value of parameters from production to optimization, collect the result,and switch back to production value. The DOE data process, which involves the statistical analysis,is running on the teach pendant. And they exchanged the experimental data and parameter optimi-zation result via the “RAPID Program and Teach Pendant Interface Program” which is developedbased on ABB “Robot Application Builder (RAB)”. RAB allows the FlexPendant to communicatewith IRC5 robot controller and exchange the data between the RAPID program and C# program.This software structure ensures that the access to the result of optimization experimental trials, theanalysis of the result, and the update of the production parameters do not disrupt the productionprogram. Moreover, the optimization tool can perform multiple parameters and multiple level DOEoptimization on multiple assembly stages simultaneously since the DOE experimental data collec-tion and DOE experiment data processing are separated.

The factory acceptance test on robotic force control torque converter assembly system wasperformed at the Ford Powertrain Assembly plant. The 3-level 5-factor 243 (35) trials DOE


of 47

optimization experiment was conducted on a three-stage assembly (the pump impeller feature thatdrives the transmission pump as the engine turns). Each trial has 10 repetitions to reduce the randomnoise in the test result. Totally 2,430 tests data are collected and analyzed from one optimizationiteration. Table 2 shows the experimental results.

In the table, Mean is mean of the assembly cycle time; StdDev is the standard deviation of theassembly cycle time; and FTT is the first time through rate (success rate). The experimental resultsshow that the average assembly cycle time is improved from 5 s (initial) to 4.6 s (third optimizationiteration). And success rate is improved from 26.6 % to 50 %. Also since in the third optimizationiteration, the verification result does not have significant improvement compared with the thirdbaseline result; therefore, the optimization is complete. The optimization tool updated the produc-tion force control assembly parameters and switched the system from optimization mode toproduction mode. This experimental result from the factory test proofs that the effectiveness ofthe proposed online DOE-based parameter optimization and the productivity of the robotic forcecontrol assembly unit is improved.

Fig. 50 The online (in production) DOE-based parameter optimization process


of 47

SummaryThis section introduces the force control robotic assembly and gives an example of torque converterassembly. DOE-based assembly parameter optimization is described. The online (in production)DOE-based parameter optimization is proposed to adapt to the manufacturing environment changes.The challenges and solutions are discussed. Finally the method is implemented and tested in forcecontrol torque converter assembly in factory.

Fig. 51 The software structure diagram of the in production DOE-based parameter optimization tool

Table 2 The in-production DOE-based parameter optimization experimental result

Mean (s) StdDev (s) FTT (%)

Baseline (initial) 5.005 2.014 26.6

Baseline (third iteration) 4.616 2.103 53.3

Verification 4.639 2.245 50


of 47

Conclusion

This chapter describes the robotic assembly technology, including force control assembly, assemblywithout force control, vision and force control-integrated assembly, and assembly on a movingproduction line. Real industrial applications are used to demonstrate how to implement thesetechniques in manufacturing automation. Because there are different process parameters that areinvolved in robotic force control assembly, parameter optimization methods are also presented. Thesteps about parameter optimization are illustrated by using a torque converter assembly process.Because the techniques are demonstrated using industrial examples, they are somewhat ready to beimplemented in production.

References

ABB Inc., (2010) Robotics application manual – force control for assembly. ABB AutomationTechnologies BA, Robotics, SE-721 68, Vasteras

Baeten J, Schutter JD (2002) Hybrid vision/force control at corners in planar robotic-contourfollowing. IEEE/ASME Trans Mechatronics 7(2):143–151

Baeten J, Bruyninckx H, Schutter JD (2002) Tool/camera configurations for eye-in-hand hybridvision/force control. In: IEEE international conference on robotics and automation, Washington,DC, pp 1704–1709

Chang W, Wu C (2002) Integrated vision and force control of a 3-D OF planar surface. In: IEEEinternational conference on control applications, Glasgow, Scotland, pp 748–753

Chen H, Zhang G, Zhang H, Fuhlbrigge T (2007) Integrated robotic system for high precisionassembly in a semi-structured environment. Assem Autom 27(3):247–252

Cho C, Kang S, Kim M, Song J-B (2005) Macro-micro manipulation with visual tracking and itsapplication to wheel assembly. Int J Control Autom Syst 3(3):461–468

Glosser GD, Newman WS (1994) The implementation of a natural admittance controller on anindustrial manipulator. In: Proceedings of the international conference on the IEEE robotics andautomation, vol 2. San Diego, CA, pp 1209–1215

Gottschlich SN, Kak AC (1989) A dynamic approach to high-precision parts mating. IEEE TransSyst Man Cybern 19(4):797–810

Gravel D (2007) Efficient method for optimization force controlled robotic assembly parameters.Presented at control and applications, 30 May–1 June

Gravel D, Newman WS (2001) Flexible robotic assembly efforts at Ford Motor Company. In:Proceedings of the 2001 I.E. international symposium on intelligent control, pp 173–182

Gravel D et al (2009) Objective metric study for DOE-based parameter optimization in robotictorque converter assembly. In: Proceeding of IEEE/RSJ international conference on intelligentrobots and systems, St. Louis, MO, pp 3832–3827

Jörg S, Langwald J, Stelter J, Hirzinger G, Natale C (2000) Flexible robot-assembly using a multi-sensory approach. In: Proceedings of the 2000 I.E. international conference on robotics andautomation, ICRA, April 24–28, San Francisco, CA

Montgomery D (2005) Design and analysis of experiments, 6th edn. Wiley, HobokenMorris DM, Haber R, Newman MS (2001) Force guided assemblies using a novel parallel manip-

ulator. In: Proceedings of IEEE international conference on robotics and automation, vol 1,Florida, USA, pp. 325–330


of 47

Morrow JD, Nelson BJ, Khosla PK (1995) Vision and force driven sensorimotor primitives forrobotic assembly skills. In: Proceedings of the IEEE international conference on intelligent robotsand systems (IROS-95)’, vol 3. Pittsburg, PA, pp 234–240

Nelson B,Morrow J, Khosla P (1995) Improved force control through visual servoing. In: Americancontrol conference, Seattle, Washington

Newman WS, Zhao Y, Pao YH (2001) Interpretation of force and moment signals for compliantpeg-in-hole assembly. Proc IEEE Int Conf Robot Autom ICRA 1:571–576

Olsson T, Bengtsson J, Johansson R, Malm H (2002) Force control and visual servoing using planarsurface identification. In: IEEE international conference on robotics and automation, WashingtonDC, pp 4211–4216

Shin IS, Nam SH, Yu HG, Roberts RG, Moon SB (2006) Conveyor visual tracking using robotvision. In: 2006 Florida conference on recent advances in robotics, FCRAR 2006, Miami, Florida

Siciliano B, Villani L (1999) Robot force control. Springer, New YorkTurges RH, Laowattana S (1994) Passive assembly of non-axisymmetric rigid parts. In: Proceedings

of the international conference on intelligent robots and systems 1994, Munich,vol 2, pp 1218–1225

voii Collani Y, Scheering C, Zhang J, Knoll A (1999) A neuro-fuzzy solution for integrated visualand force control. In: Proceeding of the IEEE international conference on multisensor fusion andintegration for intelligent systems, Taipei, pp 135–140

Wang J et al (2008) Force control new technologies for new robotics applications. Presented at the2008 I.E. international conference on technologies for practical robot applications, Woburn,10–11 Nov 2008

Xiao D, Ghosh BK, Xi N, Tarn TJ (2000) Sensor-based hybrid position/force control of a robotmanipulator in an uncelebrated environment. IEEE Trans Control Syst Technol 8(4):635–645

Xu Y, Paul RP (1990) A robot compliant wrist system for automated assembly. Proc IEEE Int ConfRobot Autom 3:1750–1755

Yoon Y, Park J, Kak AC (2006) A heterogeneous distributed visual servoing system for real-timerobotic assembly applications. In: Proceedings of the international conference on robotics andautomation, Orlando, Florida

Zhang H, Gan Z, Brogardh T, Wang J, Isaksson M (2004) Robotics technology in automotivepowertrain assembly. ABB Review, Vol. 1, pp. 13–16

Zhang G et al (2007) Automated engine piston installation using industrial robots. Presented at theICMA 2007, Harbin, 5–8 Aug 2007


of 47

Index Terms:

Assembly automation 4, 18Force control 2, 5–7, 18, 21–23, 25, 28, 31–32, 42Industrial robot 2–3, 7, 11, 18Parameter optimization 32–33, 35–36, 38, 40–42, 44Sensor fusion 18, 24, 26


of 47

robotic assembly - springer · assembling experience tells that adding hopping motion along the...

Documents