Neural Network-Based Controller with Deadzone Compensation and Estimation in a Robotic Cable-Driven Ankle Exoskeleton for Walking

Abstract

Cable-driven exoskeletons have been used as training tools and to augment muscle effort during treadmill walking. However, technical challenges remain to achieve precise control of joints using such devices. Cable tension is controlled by an electric motor to rotate a joint; however, the mapping from the input current to applied joint torque is uncertain, nonlinear, and experiences a deadzone (DZ) due to cable slackness. Further, the cable-driven device needs to account for the nonlinear, non-parameterizable uncertainty in the target muscle-tendon complex. In this paper, a neural network (NN)-based controller is developed for a robotic cable-driven ankle-foot orthosis that targets lower-leg muscles while walking. First, a NN estimator and compensator are developed to mitigate the uncertain, nonlinear DZ that exists between the motor control input and the torque applied about the ankle. The motivation for the NNs is to inject a DZ-free controller in the closed-loop error system by generating a bounded estimation of an unknown DZ preinverse function. Then, robust control terms are designed along with an additional NN that exploits the desired kinematic trajectories to mitigate uncertainty in the dynamic model. A Lyapunov-based stability analysis is developed leveraging a corollary for non-smooth systems to establish an asymptotic kinematic tracking result.