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

Area of Science:

• Robotics in neurorehabilitation
• Human-computer interaction
• Machine learning for motor control

Background:

• Current robotic neurorehabilitation often uses simple, standardized exercises.
• There's a need for robots to support complex, personalized movements for Activities of Daily Living (ADLs).
• This requires methods for clinicians to teach robots new exercises, leading to Learning by Demonstration (LfD) algorithms.

Purpose of the Study:

• To compare Task-Parameterized Gaussian Mixture Models (TPGMM) and Dynamic Movement Primitives (DMP) for generalizing upper-limb movements for ADLs.
• To extend the superior algorithm (TPGMM) to encode movements from both healthy and post-stroke individuals performing a drinking task.

Main Methods:

• Comparison of TPGMM and DMP algorithms for movement generalization.
• Implementation of TPGMM to encode upper-limb movements for a drinking task.
• Data collection from healthy participants and individuals post-stroke.

Main Results:

• TPGMM exhibited better generalization of healthy human movements across increasing task and environmental complexity than DMP and a model-based approach.
• TPGMM effectively encoded movements of individuals post-stroke, showing distinct patterns compared to healthy individuals.

Conclusions:

• TPGMM is a more effective algorithm for generalizing upper-limb movements in neurorehabilitation compared to DMP.
• TPGMM can successfully encode and differentiate movements from both healthy and post-stroke individuals, paving the way for personalized robotic rehabilitation.