Chuang-Hue Moh Spring 2002 6.836 Embodied Intelligence: Final Project.

Chuang-Hue MohSpring 2002

6.836 Embodied Intelligence: Final Project

Evolution in the Micro-Sense: An Autonomous Learning Robot

Chuang-Hue Moh

6.836 Embodied Intelligence, Spring 2002

Goal

Build a real physical robot with simple behavior and controls.

Provide the robot with simple learning capabilities and allow them the interact using subsumption.

Explore into applying genetic algorithms to the robot’s controller as a form of learning.

Complex emergent behaviors of the honeybee colony are results of interaction of individuals with simple behaviors and learning capabilities [Capaldi et. al. Ontogeny of orientation flight in honeybee revealed by harmonic radar]

Robot Design

Subsumption network architecture Exploration mode when energy is high,

recharging mode (seeks light source) when energy is low

Learns: Avoid obstacles (online self-adaptation) (current

status: completed) Navigate towards light (remembers

experiences) (current status: completed)

Experimented with genetic algorithms in an attempt to evolve a controller to avoid obstacles (current status: implemented but no experimental results yet…)

0

50

100

150

200

250

300

350

400

450

Explore

Recharge Recharge

Explore

Right Motor

Left Motor

Move Forward

Turn Right

Turn Left

s

sRandom Number

Exploress

Recharge

Light Sensor

Energy Level

Subsumption Architecture

Collision Detect

s

s

Proximity Sensor

Collision Resolve

Left Bumper Sensor

Right Bumper Sensor

ss

Robot Implementation

Lego RCXtm Microcomputer Hitachi H8/3292 micro-controller (16 MHz) with

16 KB ROM and 16 KB RAM. In-built 10-bit ADC Memory-mapped I/O 3 input / 3 output ports IR transmitter / receiver

Robot Implementation

1 x proximity sensor (light sensor + IR transmitter)

1 x light sensor (shared with proximity sensor)

2 x touch sensors (switches) 2 x 9V DC motors

Light Seeking Behavior

Remembering light intensity - simplified “eligibility trace” type data structure

Zeroing into light source location – reduce angle of search at each forward step

Dynamic lighting conditions – remembers last two light intensity levels

Demonstration

Demo available at http://www.pmg.lcs.mit.edu/~chmoh/demo.avi

Conclusion

Lessons learnt: Physical robots + real world environment simulation Too many concurrent tasks causes problems – complexity, time-slicing /

polling Sensors does not always work as expected Non-uniformity of robot movement (due to battery levels / motors) Too much abstraction is not good for robot (real-time) control

Future work: Energy level = real battery level (robot action dependent on battery level) Emergent behavior of multiple robots Learning algorithm optimization More efficient genetic algorithm