Adaptive Motion Imitation for Robot Assisted Physiotherapy Using Dynamic Time Warping and Recurrent Neural Network

Robot assisted physiotherapy has the potential to reduce the work-load of healthcare professionals and deliver interventions in the comfort of the home. In general, physiotherapy involves repeating a specific set of motions until an efficiency metric is reached. In robot assisted physiotherapy sessions, two major questions arise: 1) How to accurately quantify the similarity of the motion between the robot and the subject; 2) How to adapt the robot's behavior according to the subject's ability. In this paper, we address these two questions by proposing a new modular framework for adaptive motion imitation (AMI) using a deep long-short term recurrent neural network (LSTM-RNN) and segment online dynamic time warping (SODTW). Our framework uses the SODTW cost as a metric for quantifying the similarity between the motion of robot and subject. The LSTM-RNN takes the range of motion and the fundamental discrete Fourier transform (DFT) coefficients as inputs and uses them to predict a dynamic and periodic reference trajectory for the robot. By modifying the DFT coefficients based on the SODTW cost of the subject, the output of the LSTM-RNN is then adapted according to the imitation ability of subjects. The separation between the prediction and adaptation portions of our framework greatly simplifies testing, coding and improves the algorithm scalability. We tested the proposed AMI framework with 10 participants to verify its effectiveness. The results demonstrate the validity of the proposed framework in adapting the behavior of the robot according to the subject's imitation abilities.