Planning Behaviours

The Brain is a slow computer yet humans can skilfully play games where very fast reactions are required. A review of data on motor and cognitive effects of cerebellar lesions leads to propose that the brain minimises reaction time during skilled behaviour by eliminating online planning. Behaviour then results from direct mapping from perceptions to actions. Its is proposed that the cerebellum i) learns sensory-motor associations planned off-line and ii) enables the set of sensory-motor associations most appropriate for a given goal. In that framework, planning defines both the stimulus and the action that will be triggered. For producing effective behaviour, such a scheme requires in principle that, for each goal, a large set of sensory motor associations are prepared to cover a wide range of possible initial conditions and variations in execution. A neural network implementation of these principles is proposed with generalization properties obviating the need for planning multiple associations. After learning only the optimal plan, the NN exhibits a field of solutions that ensures robustness against variations in initial conditions and execution noise. The proposed scheme unifies purely behaviour-based approaches to robot control and approaches based on planning using internal representations. On the one hand sensory-motor associations enable fast reactions and on the other hand, being products of planning, these associations enables flexible goal adaptation. Such a principle is demonstrated with a simulated robot arm control problem and has been applied to the control of a real autonomous wheelchair.