thesispresentation

Rapid Prototyping of mobile robotsfor rough terrain usingEvolutionary Strategies

Master Thesis

Abheek Kumar BoseMaster of Science in Autonomous Systems

19th August 2005

2

Master of Science in Autonomous Systems

Rapid Prototyping of mobile robots for rough terrain using Evolutionary Strategies

Robot development is a complex blend of ...

.... and its tough!

Science Applied Mathematics Engineering

Its even tougher when we have to develop in limited time!!!

so what can we do to create such complex systems so fast??

3



Outline

● Problem Statement & Motivation● Overview of related work● Methodology● Solution & its analysis● Experiments & Results● Conclusion & Future Directions

4



● The complexity of robotic systems cause development to be slow and expensive● Robots are mostly specialized for specific applications

● Industrial manipulators are not expected to perform housekeeping tasks!● Robot design does not follow any specific standards

As a result...● Robots from separate designers/manufacturers are not compatible● Maintenance and Service can be provided only by the developers

Why is robot development a difficult problem?

What would happen if railway tracks were different in different countries? .... our situation with robots is somewhat similar!

5



● The term “rough terrain” signifies robots capable of:● Good mobility performance over outdoor environments● Carrying heavy payloads ● Working ideally in real life applications and unpredictable conditions

● Some real-life applications include:

● Rescue robots● Exploration robots● Autonomous transports

Such robots are usually:● Sophisticated but expensive● Time consuming to design and develop● Containing complex mechanisms (suspension / steering) hence difficult to maintain● Limited in flexibility & expandability to incorporate changes or system expansion

What is the big deal about “rough terrain” robot development?

6



● Introduce a development standard● Aim to reduce design and development time and cost● Construction kits are a viable approach

● Components are building blocks which foster easy development● Flexible – various forms can be obtained by differing the component assemblies● Expandable – clearly defined interfaces makes integration easier● Versatile – variants of robots can be developed with a handful of parts

However...● Construction kits have limited robustness● Developed robots have limitations in their performance and applications● “Rough Terrain” demands are very high for construction kit approaches● Unpredictable conditions push the demands further

How can we make the development better?

The construction kit approach must be assisted!!

7



● Why do our hands have 5 fingers that too of unexplainable lengths?● Why is the thumb placed where it is?● How many different tasks can we perform with only our hands?● Can you strap on your wristwatch without using your thumb?

Evolution tends to specialize... but for a variety of tasks

Evolution is a desirable assistant since:● The components of the construction kit give evolution a variety of options● The flexibility of the robot structure allows it to be “fine tuned” by evolution● The robot morphology can be ideally determined for good performance● It just sounds cooler!

Why Evolution?

8



Related WorkThe MoRob Project (Gerecke et. al. 2003)● Construction Kit for mobile robotics● Focused mainly on education● Robust system development● Limited in robot variants

The Shrimp Rover (Estier et. al. 2000)● Excellent terrain adaptability● Promising for Planetary Missions● Low payload capacity● Limited structural flexibility

The Golem Project (Pollack et. al. 2000)● Evolution in physical robotics● Direct evolution of CAD models● Integrates Rapid Prototyping with Evolution● Limited to very basic systems

9



Previous Work

The VolksBot RT● Construction Kit for Rough Terrain Robots● Variants of robots created using same components● Possible to create both direct and indirect drives● Drive System can be encapsulated

The Universal Drive Unit● Core component of the VolksBot RT kit● Flexibility in assembly over main frame● Coupled to any motor using standard couplings

10



This thesis asserts that by following a construction kit approach coupled with evolutionary strategies, it is possible to rapidly develop robust mobile robot prototypes which maintain a high mobility even in undesirable rough terrain.

Methodology

Approach: Modeling & DesignPhase 1: Analysis and Refactoring of the VolksBot RT componentsPhase 2: Robot Platform Modeling and DesignPhase 3: Mobility Enhancement for rough terrain

Approach: Simulation & EvolutionPhase 4: Morphology AnalysisPhase 5: Physical Representation and SimulationPhase 6: Morphology Optimization by Evolution

11



Solution: Modeling & Design

Refactoring the Universal Drive Unit

Limitations of the old design:● Over-dimensioned and heavy● Limited re-usability due to welding connections● Special clamping arrangements demanded assembly skills

Features of the redesigned UDU:● Lighter components (30% lighter)● Easy connection mechanisms via keys and key-ways● More components are re-usable● Less assembly skill required● More variants of the UDU possible

12




A new variant

● Vertical Drive Unit ● Two types of UDUs used● Good ground clearance which can be modified● Better turning abilities (no skid steering)● Entire drive unit can be electromagnetically shielded

Enhancing the Mobility – The parallel bogey mechanism

● Parallel Bogey: 4 bar link mechanism of parallel bars● Construction is inspired by the shrimp rover● 4 different UDUs differing only in the shaft● Full drive transmission is encapsulated within the unit● Allows a very good terrain adaptability for the robot

13




High mobility platforms

14




Type A● 4 driven wheels, 1 supporting castor● Better turning ● Better ground clearance● More compact

Type B● All 6 wheels are driven● More power needed for turning ● Ground clearance decreases at rear● More space for objects inside chassis

Both the robots are designed using only the components from the kit

● Simulation is realised through Open Dynamics Engine*● Physical parameters can be easily defined● Easy integration to the Evolution System● Provision for adding intelligent behaviours in future

15



Solution: Simulation & Evolution

Simulation of Type A

Simulation of Type B

* Open Dynamics Engine – Russel Smith (www.ode.org)

Robot Modeling● Both Type A and Type B are simulated● Simulation models are based on CAD structure● Closely concurrent with the CAD specifications● Composed of modules

● Module geometries are parameterized● Parameters define lengths and positions

16




Environment Modeling● Most vital component in the simulation design● Forms the basis for the evolution● Needs to represent a worst case scenario● Must compensate for an unpredictable terrain● Should prevent robot specialization by evolution● Should allow evolution to converge to a solution

17




Environment Description● Simple blocks are determined as most difficult obstacles for robots to scale● Environment is composed of different blocks, gaps and a staircase

● Staircase represent urban rescue scenarios● Scaling stairs demand very good mobility for robots

● Unpredictability is accounted for by randomization of environment● Block lengths (b), stair lengths (r) and gaps (g1,g2) change randomly● Block heights (h) can be changed manually during evolution

● Friction parameters can also be adjusted to make ground more slippery

18




Evolution Basics● Evolution is used as a tool to improve the robot structure● Evolution is implemented by the ISEE Platform1

● ISEE runs ENS^3 Algorithm2

1 Integrated Structure Evolution Environment (ISEE) – Fraunhofer AIS, Intelligent Dynamics Team (INDY)2 Evolution of Neural Systems by Stochastic Synthesis (ENS^3) – Hülse et. al. (2004)

19




Evolution Implementation● Important robot morphology parameters are initially determined● Each individual is represented by a net● The net has as many input neurons as the parameters● The net has one output neuron representing the fitness● The synapse strengths of this net are the robot parameters● Evolution works on these nets defining the simulated robot● Evaluation is done by driving the robot through the environment● The fitness is the distance traveled by the robot in a specified time● Environment is randomized every generation to prevent specialization● Environment difficulty is low at the beginning● Block heights are raised very slowly in very small steps

Fitness of the robot is the distance traveled in the environment

20




Evolution of Type A Robot● Evolved Parameters:

1.Body Length (Lm)2.Parallel Bogey Lever Length (l)3.Parallel Bogey Position (a)4.Lever Displacement (g)5.Leg Length (h)6.Laptop and Battery Position7.Castor Position (t)

Type A Robot: Before (left) and after (right) evolution

21




Evolution Analysis for Type A Robot● Non standard fitness path● Fitness fluctuates due to randomness● Maintains more or less a constant success:failure ratio even with increasing difficulty● High successes initially (low difficulty)

22




Evolution Analysis for Type A Robot● Parameters are selected considering the best fitness and best age of the individuals● Parameters 2, 4 and 5 are clustered

● All define the parallel bogey● Restriction signifies importance

23




Evolution of Type B Robot● Parallel Bogey dimensions remain from Type A

● Compact● Good Climbing abilities for Type A

● Evolved Parameters:1.Body Length (Lm)2.Parallel Bogey Position (a)3.Lever Displacement (g)4.Laptop Position 5.Battery Position

Type B Robot: Before (left) and after (right) evolution

24




Evolution Analysis for Type B Robot● Higher success rates than Type A● Overall fitness is higher

25




Evolution Analysis for Type B Robot● Parameters are selected considering the best fitness and best age of the individuals● Parameter 3 is clustered

● Represents a restriction on the lever displacement

26




Performance Evaluation 1 -->

● Running evolved the robots in environment● Randomization is removed● Environment difficulty increases every test cycle● All possible environments are tried out

<-- Performance Evaluation 2

● Tests for performance in uncertain conditions● Test ground square random step field● Square blocks of random heights are generated● Consequent generated gaps can trap wheels● This environment was never present during evolution

27



Experiments: Rapid Prototyping

The “People Mover Project”● Based on the 3 wheeled variant of the VolksBot RT● Intelligent Vehicle Demonstrator● Total Assembly time of approx. 7 hrs● Software development with reusable codes● Fully functioning software in approx. 2 hrs● Presented in the German Open in Paderborn, 2005

The VolksBot eXtreme Terrain● Developed directly from the Type B robot model● Component assembly by direct application of the evolved parameters● Full construction in approx. 27 hours

28



Experiments: Performance Demonstration

The VolksBot eXtreme Terrain

29



Conclusion & Future Directions

Contributions● A standard for robot development● A construction kit for real world robotics● Evolution as a viable tool for optimization

Lessons Learned● Standardized components impel rapid prototyping● Concurrency between simulation and modeling is vital● Evaluation technique is critical aspect for evolution

Future Directions● More versatile components for the construction kit● Analysis and optimization with payloads● Rough Terrain Autonomy

30



THANK YOU(for staying awake all along!)

thesispresentation

Documents

expensive robots

robots capable

term rough terrain

development time

autonomous systems19th

development standard

robot design

limited time