The Reinforcement Learning for Adaptive Prosthetics project is a collaboration with the AICML, the Glenrose Rehabilitation Hospital, and the Composite & Biomedical Materials Research Group (MechanicalEngineering, U of A). Its goal is to develop reinforcement learning (RL) algorithms that can help increase limb-deficient patients’ ability to customize and control their new prosthetic devices, while at the same time removing the need for frequent manual adjustments by patients and physiotherapists. Specifically, the project addresses the following open research questions:
- How best to translate muscle signals into control commands for a multi-function mechanical limb
- How to automatically adapt control systems to the needs of individual patients
- How to improve limb control based on patient feedback
When complete, the developed methods will increase the speed and success with which new amputees can adapt to their powered prosthetics, directly improving the quality of life for limb-deficient patients.
Since its inception in 2010, the project has demonstrated a successful first application of new actor-critic RL techniques to the domain of upper-arm myoelectric prostheses. This is an important incremental step toward the goal of fully adaptable intelligent prosthetic devices. Current work is focused on the application of new RL prediction and knowledge representation techniques to further increase the adaptability and learning potential of intelligent prostheses.