Personalized Adaptive Assistance With Reinforcement Learning Control Enhances Engagement, Performance, and Retention...

Area of Science:

• Robotics
• Rehabilitation Engineering
• Machine Learning

Background:

• Stroke survivors often require upper-limb rehabilitation to regain motor function.
• Traditional robot-assisted rehabilitation may not optimally adapt to individual patient progress.
• Personalized control strategies are crucial for effective physical human-robot interaction (pHRI).

Purpose of the Study:

• To introduce and evaluate a new Reinforcement Learning Assist-as-Needed (RL-AAN) controller.
• To compare the RL-AAN controller's performance against a conventional Iterative Learning Control Assist-as-Needed (ILC-AAN) controller.
• To assess the RL-AAN controller's ability to personalize robot assistance in real-time.

Main Methods:

• Implementation of a modified action-dependent heuristic dynamic programming (ADHDP) framework for the RL-AAN controller.
• Validation on a cable-driven, end-effector rehabilitation robot.
• Perturbation-based reaching tasks performed by healthy individuals to compare RL-AAN with ILC-AAN.

Main Results:

• The RL-AAN controller significantly reduced the required robot assistance during training.
• User active participation and task performance were promoted by the RL-AAN controller.
• Retention tests demonstrated more accurate arm-reaching trajectories after RL-AAN training compared to ILC-AAN training.

Conclusions:

• The RL-AAN controller shows potential for enhancing exercise-based rehabilitation by personalizing robot assistance.
• This approach facilitates greater user engagement and improved motor recovery outcomes.
• The study contributes to the development of adaptive control strategies for physical human-robot interaction in rehabilitation.