Solving credit assignment problem in behavior coordination learning via robot action decomposition

In behavior coordination, several primitive behaviors are “combined” to generate a resultant action to drive the robot. The weights across the primitive behaviors should be properly determined according to the situations that the robot encounters in order to successfully avoid collisions with obstacles and accomplish the assigned task. Behavior coordination learning is proposed to learn the mapping between the situations encountered by the robot and the weight combinations on primitive behaviors from observed resultant behavior of the robot. The paper proposes an action decomposition algorithm to automatically derive the weights across primitive behaviors from an observed resultant behavior with minimum weight variations along time by a local optimization scheme. Several examples on simulated and experimental data are presented to demonstrate the computation of action decomposition