Offline-to-Online Learning Enabled Robust Control for Uncertain Robotic Systems Pursuing Constraint-Following

Abstract

A major challenge in robust control design of robotic systems is finding a comprehensive uncertainty bound (CUB) with low conservativeness for uncertainty compensation. This study proposes a two-phase learning approach to learn the CUB for robust control of robotic systems, considering uncertainties with unknown bounds. The goal is to drive the system to follow a class of servo constraints that may be nonholonomic, i.e., constraint-following control (CFC) design. The first phase trains a deep neural network (DNN) to approximate the ensemble system uncertainty using offline supervised learning. The second phase constructs an adaptive law to learn the CUB online, covering the offline learning error. To our knowledge, this is the first CFC combining DNN with adaptive law to learn a less conservative CUB to save control effort and to eliminate complex manual derivations. The effectiveness and merits of the proposed control are endorsed by theoretical proofs, simulations, and experiments on a quadrotor unmanned aerial vehicle.