Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction

A research team from institutions including Northwestern University and the Shirley Ryan AbilityLab has published a novel approach to automating post-stroke rehabilitation therapy. Their paper, 'Learning Therapist Policy from Therapist-Exoskeleton-Patient Interaction,' introduces two key innovations: a Patient-Therapist Force Field (PTFF) for visualizing interaction dynamics and a Synthetic Therapist (ST) machine learning model. The ST is designed to learn and replicate a human therapist's physical responses during robot-assisted gait training, aiming to reduce the physical strain on therapists and potentially increase therapy consistency and intensity over long-term treatment.

The technical core uses a Variational Autoencoder (VAE) to compress patient and therapist stride kinematics into a shared latent space, modeling their interaction with a Gaussian Mixture Model (GMM). The predictive ST model itself is a Long Short-Term Memory (LSTM) network trained via leave-one-out cross-validation on data from eight patients. It predicts the joint torques a therapist would apply based solely on real-time patient kinematics. This model was integrated into a Robot Operating System (ROS)-based exoskeleton controller for preliminary testing. The system's promise lies in its data-driven approach—learning directly from expert human demonstrations—which could lead to more adaptive and personalized robotic therapy, allowing the human clinician to shift from physically demanding assistance to higher-level supervision and strategy.