Learning Control of Compliant Robots Using Singular Perturbation

Rehabilitation exoskeletons can effectively assist the elderly and patients with nerve loss in their activities of daily living, but traditional rigid rehabilitation exoskeletons may impede the normal movement of the body’s joints and add additional motion constraints, which may result in injury to the wearer.With the introduction of serial elastic actuators (SEA), the flexible rehabilitation exoskeleton has a more natural human-machine interaction and can improve the wearer’s safety and mobility during rehabilitation.However, the introduction of SEA greatly increases the complexity of the control system. The paper presents a learning control strategy for a rehabilitation exoskeleton robot with serial elastic actuators and modeling uncertainty. Using singular perturbation, this controller converts the fourth-order underperturbed dynamics of the robot into a second-order slow-varying subsystem and a second-order fast-varying subsystem. The fast-varying subsystem is then stabilized using a derivative controller. To achieve an efficient estimation of the modeling uncertainty, and to make this subsystem globally asymptotically stable, a concurrent learning (CL) approach is used to design a concurrent learning controller for the slow-varying subsystem. The control stability is analyzed using Hoppenstead’s theorem and the Krasovskii-LaSalle invariance principle, and the control effectiveness is validated by simulations.