Can Learning From Demonstration Approaches Encode and Generalise Human Movements for Neurorehabilitation?

The use of robotic systems in Upper-Limb (UL) neurorehabilitation typically involves semi-standardised, simple movement exercises controlled by the robot. However, alternative approaches aim to support more complex movements that align with Activities of Daily Living (ADLs) and offer greater customisation of interactions tailored to individual patients by clinicians. These approaches, however, require increased therapist involvement, which underscores the need for methods that allow clinicians to teach a set of exercises to the robot. This has led to the development of various Learning by Demonstration (LfD) algorithms. These algorithms require the ability to encode the movements demonstrated by the clinician and generalise them across various task variations. Towards this goal, this study compares two existing LfD algorithms, Task-Parameterised Gaussian mixture models (TPGMM) and Dynamic Movement Primitives (DMP), in their capacity to generalise UL movements required for ADLs. The study then extends the best performing algorithm - TPGMM - to encode movements from both healthy and poststroke participants performing a drinking task. TPGMM shows better performance in generalising healthy human movements for tasks and environments of increasing complexity when compared to a model-based approach and DMPs. TPGMM further shows that it better encodes movements of individuals post-stroke compared to ones of healthy individuals.