Experiments in robot learning

Abstract

This chapter reviews two robots that can learn how the world behaves in, and improves their own performance over time based on the information gathered. The underlying architecture of such robots would be of the abstract agent variety and implemented in CommonLisp. The abstract agent defines a framework for defining and testing different learning agents. Also, the architecture of the abstract agent allows a new agent to be created by plugging in a new module or recombining old modules. The experimental task domain defined for the robots is the operation of a tilting tray. A robotic manipulator holds a square tray, bounded by four vertical walls, which may be tilted at any angle desired. An object, typically a square block, slides freely in the tray. By a judicious choice of tilting sequence, the robot is able to position the object in the tray. The chief complication of the tilting-tray occurs when the object hits a wall. Now, one of the experimental robots is proposed to have been fitted with the rote agent performs rote-learning of tilt motions mechanics while the other with the inductive agent, which does a simple version-space learning to generalize in the space of tilt-angles. Both agents use exhaustive search to construct plans. When no such plan can be constructed, a single random tilt is chosen.