Adaptive Modeling and Control of an Upper-Limb Rehabilitation Robot Using RBF Neural Networks

Abstract

Robot-assisted rehabilitation following neurological injury is most successful when subject participation is maximized in the training tasks. Developing control strategies that can provide subject-specific assistance is accordingly an active area of research. For robot-assisted rehabilitation training, it is challenging to adapt the robotic assistance to each patient’s impairment, and model-based control methods in previous studies are difficult to implement because of the computational complexity of human-robot interaction dynamics and changes of human active efforts during rehabilitation exercises. This study implements adaptive modeling and control for an two-DOF upper-limb rehabilitation robot by combining an RBF-based feedforward controller with a feedback impedance controller. Simulation and experiment results show that, the RBF neural network is able to adaptively establish the human-robot dynamics as well as estimating the human efforts, and the impedance controller guarantees compliant human-robot interaction and regulates the maximal tolerated tracking error. Besides, the proposed controller is defined in the robot workspace, thus is easy to be generalized to be used for multi-DOFs exoskeleton-type rehabilitation robots.