Unsupervised learning of sensory-motor primitives

The search for motor primitives has captured the attention of researches in both biological and computational motor control. Yet a theory of how to construct such primitives from first principles is lacking. Here we propose to do that by building a compact forward model of the sensory-motor periphery via unsupervised learning. We also propose a method for probabilistic inversion of the forward model, which yields low-level feedback loops that can simplify control. The idea is applied to simulated biomechanical systems of varying levels of detail.