Learning to coordinate behaviors for robot navigation

Abstract

Past approaches to navigating behavior-based robots have relied on a priori determined arbitration schemes to control the total behavior of a robot. Such arbitrators can only resolve deadlocks between competing behaviors by introducing randomness. Without monitoring its self-behavior and using state information a robot cannot improve its performance. We show that by continually monitoring its actions (the output of behaviors) a robot can discover the deadlocking features (local minima) in the environment that cause failure. The robot can determine the correct arbitration sequence between its behaviors that is needed to resolve the conflict. We present experimental results of a Nomad robot discovering and then using environment features to escape local minima while navigating in unknown environments.