Online Learning of the Dynamical Internal Model of Transfemoral Prosthesis for Enhancing Compliance

Powered prostheses hold promises to restore adaptive and robust locomotion to lower-limb amputees. However, their daily use is still challenged by several shortcomings, on top of which those related to their control methods. This letter reports the development of an adaptive controller for a transfemoral prosthesis that combines a predictive torque component with a feedback error-correction mechanism. The predictive module is based on the Locally Weighted Projection Regression (LWPR) algorithm that achieves nonlinear function approximation of a dynamical model of the prosthesis joints. The performance of the proposed control strategy are assessed with a simulated biped walker with a unilateral transfemoral amputation. Results show that the LWPR-based module provides accurate predictions of the ankle and knee torques, resulting in a precise position tracking. This allows reducing the gains of the feedback error-correction mechanism by one order of magnitude, leading to a feedback contribution to the total joint torque lower than 3% and 8% for the ankle and the knee, respectively. Compliance of both prosthesis joints is enhanced accordingly. In addition, the control architecture is robust to speed changes while the joint dynamical internal model is continuously learned. This approach is thus promising for the development of adaptive controllers for lower-limb prostheses.