Toward Physician-Level Performance in Robot-Assisted Ankle Rehabilitation via Imitation Learning with Empirical and Temporal Adaptation

Robot-assisted ankle rehabilitation training imitating physician's professional techniques is highly important for promoting personalized training and improving clinical outcomes. In this work, we propose a two-level kernelized movement primitives (2-level-KMP) imitation learning algorithm under the kernelized movement primitives (KMP) framework, which reproduces physician's experience and optimizes the imitation trajectory during rehabilitation, to realize physician-level performance in robot-assisted ankle rehabilitation training. First, a KMP process combined with a Bayesian optimizer is used to imitate the rehabilitation trajectory. Second, the other KMP process is used to smooth the imitation trajectory further. Then the two KMP processes combined with patient-in-the-loop optimization (PILO) realize temporal rehabilitation adaptation. Finally, the 2-level-KMP algorithm is reproduced on a parallel ankle rehabilitation robot (PARR), which enables the patient's passive rehabilitation training to be empirical and adaptive. Ten ankle dysfunction patients were involved in clinical experiments, with the results showing that the proposed algorithm can accurately reproduce physician's trajectories and modulate trajectories based on patient's feedback. After ten rehabilitation exercises, the number of modulation points calculated from patient's torque feedback decreases by 85.19% on average compared with the beginning stage. A comparison between the 2-level KMP algorithm and existing algorithms shows that the 2-level-KMP algorithm can better ensure smoothness and retain the shape of the trajectory during trajectory modulation, ensuring the safety of ankle rehabilitation and retaining the experience of the physician.