On-line Learning of an Object Manipulation Behavior for Legged Robots

In this work we address on-line learning of behaviors for legged robots. We present a m ethodology for generating and adapting object manipulation behaviors using reinforcement learning. An AIBO four-legged robot carries cylinders to a desired position by pushing them with its frontal legs. The movement of a cylinder depends on where the robot exerts forces on it, and upon several physical characteristics (geometry, weight, floor roughness, etc.). Initially the robot doesn't know this dynamics and it learns it from its own experience. By continuously evaluating the action that minimizes the ti me to achieve the goal, and using a global configuration fitness function, the robots can complete the target configuration, and learn the tradeoff between pushing a given cylinder on a given position and moving itself to a new pushing position. Successful results on the application of this methodology are presented.