DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 1

Adaptive Intelligent Mobile Robotics

Leslie Pack Kaelbling

Artificial Intelligence Laboratory

MIT

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 2

Progress to Date

• Erik the Red• Video game environment• Optical flow implementation• Fast bootstrapped reinforcement learning

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 3

Erik the Red

RWI B21 robot• camera, sonars, laser range-finder, infrareds• 3 Linux machines• ported our framework for writing debuggable code

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 4

Erik the Red

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 5

Crystal Space

Public-domain video-game environment• complex graphics• other agents• highly modifiable

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 6

Crystal Space

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 7

Optical Flow

Get range information visually by computing optical flow field

• nearer objects cause flow of higher magnitude• expansion pattern means you’re going to hit• rate of expansion tells you when• elegant control laws based on center and rate of

expansion (derived from human and fly behavior)

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 8

Optical Flow in Crystal Space

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 9

Making RL Really Work

Typical RL methods require far too much data to be practical in an online setting. Address the problem by

• strong generalization techniques• using human input to bootstrap

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 10

JAQL

Learning a value function in a continuous state and action space

• based on locally weighted regression (fancy version of nearest neighbor)

• algorithm knows what it knows• use meta-knowledge to be conservative about

dynamic-programming updates

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 11

Incorporating Human Input

Humans can help a lot, even if they can’t perform the task very well.

• Provide some initial successful trajectories through the space

• Trajectories are not used for supervised learning, but to guide the reinforcement-learning methods through useful parts of the space

• Learn models of the dynamics of the world and of the reward structure

• Once learned models are good, use them to update the value function and policy as well.

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 12

Simple Experiment

The “hill-car” problem in two continuous dimensions• Regular RL methods take thousands of trials to

learn a reasonable policy• JAQL takes 11 inefficient but eventually successful

trails generated by humans to get 80% performance

• 10 more subsequent trials generate high quality performance in the whole space

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 13

Success Percentage

0

10

20

30

40

50

60

70

80

90

100

subsequent training runs

JAQLQL

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 14

Trial Length (200 max)

0

20

40

60

80

100

120

140

160

180

subsequent training runs

JAQLQL

54-step

optimum

DARPA Mobile Autonomous Robot SoftwareLeslie Pack Kaelbling; January 2000 15

Next Steps

• Implement optical-flow control algorithms on robot• Apply RL techniques to tune parameters in control

algorithms on robot in real time• corridor following using sonar and laser• obstacle avoidance using optical flow

• Build highly complex simulated environment• Integrate planning and learning in multi-layer

system