Emergent Control and Planning in an Autonomous Vehicle

Abstract

We use a connectionist network trained with rein-\nforcement to control both an autonomous robot ve-\nhicle and a simulated robot. We show that given\nappropriate sensory data and architectural struc-\nture, a network can learn to control the robot for\na simple navigation problem. We then investigate a\nmore complex goal-based problem and examine the\nplan-like behavior that emerges.