6.836 Embodied Intelligence Project
Evolution in the Micro-Sense: An Autonomous Learning Robot

Abstract

Autonomous learning robots have the advantage over manually programmed robots in that they are able to adapt to varying internal conditions of the robot (e.g., energy levels), as well as to dynamic external environmental conditions (e.g., light, friction). In this project, we implemented a robot that has the ability to learn obstacle avoidance using online self-adaptation. We enhanced the robot's learning ability by incorporating a light seeking behavior in the robot when its internal energy level is low. This light seeking behavior uses a tree-like data structure that is based on the concept of eligibility traces, which is commonly used in Q-learning algorithms (reinforcement learning). We also implemented a genetic algorithm on the robot and experimented with using genetic algorithm as a form of robot learning. The robot was built using the Lego RCX microcomputer and was based on Rodney Brook's subsumption architecture.

Copyright 2000 © Chuang-Hue Moh, MIT