Online Human Behavior Learning via Dynamic Regressor Extension and Mixing With Fixed-Time Convergence

To improve the hybrid augmented intelligence of a human-in-the-loop (HiTL) control system, it is desirable to investigate the issue of human behavior learning (HBL), i.e., empower the machine to understand how a human expert performs a manipulation task. The human expert is commonly modeled as an optimal controller with unknown weighting matrices that depict the tradeoff between different control objectives. Therefore, the goal of this article is to determine the weighting matrices of the human objective function with fast convergence rate, which is usually pursued to achieve better efficiency and performance in practice. Accordingly, for a class of HiTL system, we propose a novel adaptive inverse optimal control (IOC) approach for online learning human behavior with a fixed-time guarantee. Our proposed method consists of two parts. In the first step, a dynamic regressor extension and mixing (DREM)-based estimation method is used for online learning of the human feedback gain with fixed-time convergence using the demonstrated system state measurement only. Then, with the estimated human feedback gain, a semidefinite programming (SDP) problem is solved to determine the weighting matrices of the human objective function. The simulation and the experiment on a steering control have validated the effectiveness and applicability of the developed approach.