Behavior Control for Robotic Exploration of Planetary

Surfaces

Written by Erann Gat, Rajiv Desai, Robert Ivlev, John Loch and David P Miller

Presented By Tony Morelli 9/30/2004

Abstract

● Describes robots developed at JPL (Jet Describes robots developed at JPL (Jet Propulsion Laboratory)Propulsion Laboratory)

● Demonstrate using behavior-control Demonstrate using behavior-control approach to control small robots on approach to control small robots on planetary surfacesplanetary surfaces

● Behavior-Control uses very little Behavior-Control uses very little computation.computation.

Introduction

● Cannot remote control robots from Earth because of the delay

● Size is limited by power, not physical size● 3 ways to power a robot

– Radioisotope Thermal Generators (decay of Plutonium)

– Photovoltaic cells – Require heavy batteries– Non-Rechargable batteries – Short life

Behavior control

● (Rod) Brooks – Decompose the problem by task rather than function – Subsumption

● Advantages of Behavior Based Control– Fast behaviors are not slowed down by slow

behaviors (act independent of each other)– Task Specific so designers can simplify the

behavior

ALFA – A Language For Action

● Programming Language to describe reactive behavior-control mechanisms for autonomous robots

● Consists of Modules connected by Channels– Module – Converts inputs to a set of outputs– Channel – Dataflow – Data from Modules or

sensors● Similar to Subsumption

– No Wires – Easier to add modules– Provide layers of computational abstraction

rather than layers of functionality

ALFA – Code Sample

Tooth - Overview

● 30 cm X 20 cm – Indoor Robot● 1 Bit Sensors

– Grippers and rear bumper– Infrared Proximity Sensors

● Analog Sensors– Photo Cells (Find Light Beacon)– Tachometer on the drive motors

● Used 3.5kBytes of EEPROM and 100 bytes of RAM

Tooth – Control Structure

● Drive Processor/Grasp Processor● Bottom Up Design

– Cooridinating the Drive and Steering Motors– Backing up and getting out of endless loops– Picking up/Dropping objects – Head to beacon

Tooth Control Structure

Tooth – Getting Out of Loops and Dead Ends

● Unthrash Module– Lower priority than obstacle avoidance– Counts the number of times the robot changes

direction in a certain amount of time and tunrs at a random direction if it thinks it's stuck

● Dead ends – If the Robot hits a dead end it will back up, then try to go forward. If it hits a wall again, it will back up more the next time.

● Grasp Module – If it tries too many times to pick up something, it will give up

● Forward turning radius is different than backwards

Tooth - Results

● No way of searching out objects, just finds them while wandering around

● Very Robust● Could not handle wires, holes or bright lights

Rocky III - Overview

● Demonstrate behavior control could be used in a realistic planetary mission

● Infrared beacon detector● 10 kBytes of RAM● Weighs 18kg

Rocky III – Control Structure

● 3 Layers nearly identical to Tooth– Speed and Direction– Obstacle Avoidance– Sequencer

Rocky III - Results

● Very Reliable– 90% of the time completes its mission

● First example of an autonomous that operates in outdoor natural terrain that performs both navigation and manipulation

Rocky IV - Overview

● Chasis is virtually identical to Rocky III● Weighs 7.5kg● Construction Materials were modified to work

in the climate on mars.● 1 Master Processor and 3 slave processors

Rocky IV - Status

● Not yet complete● Every aspect of a Mars mission has been

demonstrated● Hardware Issues – Activating the rock

chipper caused the computer to crash (Obviously not software related)

Discussion

● Behavior control succeeds because action selection is not a difficult problem.

● ALFA code is easy to write, debug, and re-use

● Other robots were larger because they were required to scale a 1 meter tall objects

● Few simple sensors work as well as a lot of complex sensors

Summary and Conclusion

● Low power consumption is a necessity● Low CPU usage to save power● Used a modified version of subsumption● ALFA seperates data flow computations from

state machine computations● As complex as other State of the Art robots

Questions?