Learning control of complex skills

This dissertation presents a hierarchical controller which can learn to perform complex motor skills. Humans routinely coordinate many degrees of freedom smoothly and effortlessly to achieve complex goals. Moreover, we are good at learning new patterns of coordination to produce new skills. Robots and artificial systems, on the other hand, typically have difficulty with the kinds of behaviors that come most naturally to us. Skills such as running, skiing, playing basketball, or diving involve complex nonlinear dynamics, many degrees of freedom, and behavioral goals that can be difficult to specify mathematically; goals such as “ski down the mountain without falling down” or “shoot a layup” must be translated from linguistic requirements into dynamic system constraints. The focus in this dissertation will be on the skill of platform diving, in which the diver's goal is to execute a certain dive and enter the water in a fully-extended, vertical position. Controlling a simulated diver is a difficult problem for standard control and planning algorithms; conservation of angular momentum gives the system dynamics a nonholonomic constraint with nonlinear drift. In this dissertation, ideas from the fields of biological motor control and learning are combined with new learning algorithms in the design of a hierarchical controller which learns to dive. At the lower level of the control hierarchy, each degree of freedom in the diver's joints is assigned a controller based on biological pattern generators for fast, single-joint movements. These controllers contain neural networks, which are trained on data generated by simulation. The higher level of the control hierarchy incorporates ideas from human skill learning: to achieve a desired behavior pattern, a human learning a new skill uses information from instructors and from watching other performers to build a mental model of the task requirements, and then practices to refine the parameters of this behavioral model. In the high-level controller, each dive is represented as a sequence of multi-joint synergies. The controller learns initial estimates of the timing of these synergies from observational data and then refines these estimates through Q-learning with repeated simulations.