ABSTRACT
A computational approach to motor skill learning is presented, with an emphasis on problems involving excess degrees of freedom. The approach is based on the notions of (a) an internal forward model of the environment, and (b) intrinsic constraints on motor learning. A forward model allows constraints in distal, task-based coordinate systems to be converted into constraints in motoric coordinate systems, even when the mapping between these coordinate systems is many-to-one. Task constraints are augmented by a set of internal intrinsic constraints on learning. The intrinsic constraints are a relatively small class of functional forms that are assumed to apply broadly across tasks. I discuss how this approach can be implemented using network learning algorithms and demonstrate applications of the approach to inverse kinematic learning, trajectory formation, composite feedback-feedforward control, and the learning of input parameters.
