Imitation Learning Using Gaussian Mixture Models and Dynamic Movement Primitives for Rehabilitation Exoskeletons: A Comparison

Abstract

Trajectory generation for upper-limb rehabilitation exoskeletons is crucial for post-stroke therapy success. Learning by Demonstrations (LbD) techniques have emerged as powerful tools to extract smooth and accurate human skills from physiotherapists' demonstrated movements, increasing the benefits of robotic therapy and therapists' acceptability. Nevertheless, so far no concordance on the best approach to perform LbD for rehabilitation exercises has been reached. In this work, we perform a detailed analysis to compare the performances and advantages of two well-known LbD methods: the deterministic approach called Dynamic motion primitives (DMPs) and the probabilistic Gaussian mixture models and regression method (GMM-GMR). We validated and compared the proposed approaches using multiple databases of trajectories performed by physiotherapists, generated using the AGREE exoskeleton by Politecnico di Milano. Results show that the implemented methods outperform the corresponding state-of-the-art polynomial trajectories in precision and human likeness, evaluated through metrics such as Spectral Arc Length and Logarithmic Dimensionless Jerk. Our analysis reveals that both methods have potential, but for different purposes. GMM-GMR excels in precisely reproducing therapists' movements and should be preferred when the clinician wants to enhance particular gestures. DMPs give smoother trajectories, especially when dealing with smaller datasets, and can better comply with therapy timings.