Progressive-Learning-Based Assist-as-Needed Control for Ankle Rehabilitation

This article proposes a progressive-learning-based assist-as-needed (AAN) control scheme for ankle rehabilitation. To quantify the training performance, a fuzzy logic (FL) system is established to generate a holistic metric based on multiple kinematic and dynamic indicators. Subsequently, a cost function that contains both the tracking error and robot stiffness is constructed. A novel learning scheme is then proposed to enhance subjects’ engagement, leveraging the FL metric to uphold a declining trend in the robot's stiffness. The system stability is analyzed using the Lyapunov theory, the control ultimate bounds are specified and the effects of parameter tuning are discussed. Experiments are conducted on an ankle robot and the minimal assist-as-needed (MAAN) scheme is adopted for comparison. With a training session consisting of 11 trials, the quantitative performance evaluations, individual error convergences, progressive stiffness learning and human–robot interaction are evaluated. It is shown that within eight trials under the progressive AAN and MAAN, the robot assistive torques have an average reduction of 13.45% and 20.25% while subjects’ active torques are increased by 56.53% and 58.39%, respectively. During the late stage of training, the progressive AAN further improves two criteria by 9.44% and 6.29%, while the MAAN partially loses subjects’ participation (active torques are reduced by 36.38%) due to the occurrence of motion adaption.