Toward learning to control robot behaviors

It is important to methodologically distinguish between the environment dyn amics, the agent dynamics and their coupling. In order to make this distinc tion clearer, let us define the following vocabulary. We will call a sensor y-motor loop the internal mechanism linking perception to command whether f ixed or adaptive and however sophisticated the percepts and commands can be for the sake of generality. We will call behavior the externally observed behavior tin the intuitive sense) produced by the execution of such a senso ry-motor loop in a particular context. The originality of this work is, fir st, the formulation of reinforcement learning based on the sensory-motor dy namics and not the environment dynamics, and, second, to present a flexible selection learning mechanism of several sensory-motor loops and its formul ation as a case of adaptive optimal control. Selection is presented as a wa y to coordinate independently developed sensory-motor loops.