Learn from Therapists’ Demonstrations Approaches for Robotic Rehabilitation Exercises

Rehabilitation exoskeletons face acceptability issues due to their inability to produce trajectories representing therapists’ desires. Learning by Demonstration can offer a solution to this issue by using therapists’ demonstrations to teach the robot how to perform movements. While various methods for teaching movements to a robot from human demonstrations and manipulation have been proposed, no discussion about the best one for rehabilitation purposes has been conducted. Our work aims to discuss a Hidden-Markov-Model-based (HMM) approach, integrated with a human-likeness and a joint-synchronization algorithm, and compare it with two other well-known approaches we previously validated for rehabilitation. Applying the three methods to some therapists-generated datasets, produced with the exoskeleton AGREE and representing 5 different exercises, we verified that our HMM approach is better both in terms of human likeness (as measured by the SPARC metric), and compliance with therapy time ...