Motion Alphabet Augmentation Based on Past Experiences

Abstract

Multi-modal control is a commonly used design tool for breaking up complex control tasks into sequences of simpler tasks. It has previously been shown that rapidly-exploring randomized trees (as well as other viable approaches) can be used for reachability computations given a set of modes, and reinforcement learning can be performed over the reachable set to obtain the optimal control sequence. In this paper, we investigate the problem of adding new modes to a motion description language in a structured manner. We formalize an approach for augmenting the motion alphabet by adding new modes to reduce the complexity of the control program. In particular, we show a general technique for combining recurring mode sequences into one smooth "metamode". This problem is solved using a variational approach and numerical examples illustrate the feasibility of the proposed method.