A Learning-Based Decentralized Optimal Control Method for Modular and Reconfigurable Robots with Uncertain Environment

Abstract

This paper presents a novel decentralized control approach for modular and reconfigurable robots (MRRs) with uncertain environment contact under a learning-based optimal compensation strategy. Unlike the known optimal control methods that are merely suitable for specific classes of robotic systems without implementing dynamic compensations, in this investigation, the dynamic model of the MRR system is described as a synthesis of interconnected subsystems, in which the obtainable local dynamic information is utilized effectively to construct the feedback controller, thus making the decentralized optimal control problem of the MRR system be formulated as an optimal compensation issue of the model uncertainty. A policy iteration algorithm is employed to solve the Hamilton-Jacobi-Bellman (HJB) equation with a modified cost function, which is approximated by constructing a critic neural network, and then the approximate optimal control policy can be derived. The asymptotic stability of the closed-loop MRR system is proved by using the Lyapunov theory. At last, simulations are performed to verify the effectiveness of the proposed decentralized optimal control approach.