Self-Learning Distributed MPC for Multirobot Systems with State-Dependent Uncertainties

Abstract

Distributed model predictive control (DMPC) is an effective method for constrained cooperative control of multirobot systems(MRSs). However, conventional DMPC usually relies on known system models and periodic communication, which may lead to degraded performance and unnecessary resource consumption in the presence of unknown state-dependent uncertainty. To address this problem, this paper proposes a self-learning distributed MPC (SLDMPC) method for uncertain MRSs. Gaussian process regression (GPR) is employed to learn the unknown uncertainty online, and a distributed self-learning mechanism is designed to determine the next data transmission and controller update instant based on the learning error bound. In this way, unnecessary GP model updates and communication can be reduced while maintaining control performance. Sufficient conditions are established to guarantee Zeno-free behavior, recursive feasibility, and closed-loop stability. Simulation results on a MRSs verify the effectiveness of the proposed method.