Boosting Robot Learning and Control with Domain Constraints

In this short paper, we exploit robotics domain knowledge, be it acquired from humans or self-organization, to alleviate learning and control challenges from directly dealing with raw demonstrations or sparse reward signals. We take a unified latent variable perspective in incorporating domain constraints. The latent variables are regarded as task parameters or representations, which rationalize task observations with a generative model. The constraints can thus be specified with structured latent variables. Different from many related works, we explore latent structures that are computationally feasible and robotics-oriented to facilitate both task learning and control synthesis. The paper will briefly discuss adopted structures ranging from parameter dependency, modality and dynamical associativity, extending imitation learning such as inverse optimal control and deep generative models. The framework is shown to be effective in a range of manipulation tasks, including 1) learning variable impedance controllers in robotic handwriting; 2) boosting motion synthesis for writing novel symbols; 3) reasoning an internal model to score a balltarget under malfunctioning visual input.