Learning a Shape-adaptive Assist-as-needed Rehabilitation Policy from Therapist-informed Input

TL;DR

The paper addresses safe, adaptive assist-as-needed rehabilitation through a therapist-in-the-loop framework using two telerobotic collaborators. It encodes therapist corrective input as latent via-points and learns a shape-adaptive policy that partially deforms the patient reference trajectory according to patient motion preferences, while reproducing therapist input with a regression model. Experimental validation on a dual-robot telerehabilitation setup with two tasks shows reduced corrective forces and improved movement smoothness compared to baselines, and demonstrates the ability to reproduce therapist skills across sessions. The approach enables remote, therapist-guided AAN therapy with enhanced safety and active patient participation, offering a practical route toward clinical deployment and future 5G-enabled telerehabilitation.

Abstract

Therapist-in-the-loop robotic rehabilitation has shown great promise in enhancing rehabilitation outcomes by integrating the strengths of therapists and robotic systems. However, its broader adoption remains limited due to insufficient safe interaction and limited adaptation capability. This article proposes a novel telerobotics-mediated framework that enables therapists to intuitively and safely deliver assist-as-needed~(AAN) therapy based on two primary contributions. First, our framework encodes the therapist-informed corrective force into via-points in a latent space, allowing the therapist to provide only minimal assistance while encouraging patient maintaining own motion preferences. Second, a shape-adaptive ANN rehabilitation policy is learned to partially and progressively deform the reference trajectory for movement therapy based on encoded patient motion preferences and therapist-informed via-points. The effectiveness of the proposed shape-adaptive AAN strategy was validated on a telerobotic rehabilitation system using two representative tasks. The results demonstrate its practicality for remote AAN therapy and its superiority over two state-of-the-art methods in reducing corrective force and improving movement smoothness.

Figures (8)

• Figure 1: Overview of learning our shape-adaptive AAN rehabilitation policy based on two isomorphic robots interacting with patient and therapist through the end-effector, separately. The therapist first selects a specific rehabilitation goal, whose desired motion is displayed on both patient-side and therapist-side GUIs. Two impedance controllers with pretested control parameters are implemented for patient-side and therapist-side robots to repetitively complete the therapy. On the patient side, the reference trajectory is iteratively deformed during movement therapy. The patient motion preferences is encoded from previously collected actual trajectories. On the therapist side, another impedance controller is implemented to reproduce the actual motion of patient in real time, allowing the therapist to apply the corrective forces. The via-points are extracted for reference trajectory deformation, which are the transmitted to patient-side robot for the implementation of next therapy iteration.
• Figure 2: Illustration of patient motion preferences encoding, therapist-informed via-point extraction, and reference trajectory deformation in the proposed framework. (a) Gray lines are patient's actual motion trajectories. The green line indicates the estimated mean motion trajectory according to Section \ref{['subsec-motion-preference-encoding']}. Red scatter points indicate therapist-informed via-points inferred from the therapist-informed corrective force and the desired motion. The blue line indicates the generated reference trajectory according to Section \ref{['subsec-trajectory-deformation']}. (b) Gray lines represent actual motion trajectories reproduced by therapist-side robot, while therapist applies the effective corrective force as the red arrows to inform the via-points.
• Figure 3: Illustration of the artificial motor disability induced using an elastic band to provide resistance. (a) Patient at Stage #1 using a band with lower elasticity. (b) Patient at Stage #2 using a band with higher elasticity.
• Figure 4: Performance of Task #1 by the patient at Stage #1. The first row illustrates the encoded patient motion preferences and the generated reference trajectories, while the second row shows the corrective forces applied by the therapist. (a) Iteration 1; (b) Iteration 3; (c) Iteration 6; (d) Iteration 9.
• Figure 5: Performance of Task #2 by the patient at Stage #1. The first row illustrates the encoded patient motion preferences and the generated reference trajectories, while the second row shows the corrective forces applied by the therapist. (a) Iteration 1; (b) Iteration 3; (c) Iteration 6; (d) Iteration 9.