Multi-Joint Adaptive Motion Imitation in Robot-Assisted Physiotherapy with Dynamic Time Warping and Recurrent Neural Networks

Robot-assisted physiotherapy offers a promising avenue for easing the burden on healthcare professionals and providing treatment in the comfort of one's home. Typically, physiotherapy requires the repetitive movements until a certain efficiency metric is achieved. In the field of robot-assisted physiotherapy challenges include accurately determining the quality of imitation between robot and human movements, and tailoring the robot behavior to match the subject's abilities. This paper presents an innovative modular framework for Adaptive Motion Imitation (AMI) in the context of multi-joint robot-assisted physiotherapy. The proposed framework utilizes a deep Gated Recurrent Unit (GRU) Neural Network and Segment Online Dynamic Time Warping (SODTW). The SODTW cost is employed as a measure to determine the closeness between the movements of the robot and the subject. The GRU, which uses the range of motions and the fundamental frequency components of joint trajectories as inputs, forecasts dynamic and periodic reference trajectories for the robot joints. By modifying the input frequency coefficients according to the subject's SODTW cost, the output of the GRU is adapted to adapt the robot's motion with the subject's imitation capabilities. The division of the prediction and adaptation elements of our framework greatly streamlines testing and coding, and boosts the scalability of the algorithm. The efficacy of the proposed AMI framework was experimentally assessed with a group of 15 participants and the social robot Zeno in our lab. The results demonstrate the validity of the proposed framework in adapting the behavior of the robot according to the subject's imitation abilities.