Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Imitation Learning for Adaptive Control of a Virtual Soft Exoglove

Shirui Lyu, Vittorio Caggiano, Matteo Leonetti, Dario Farina, Letizia Gionfrida

TL;DR

The paper tackles personalized hand rehabilitation for motor-impaired users by combining imitation learning from healthy subject videos with a biologically accurate musculoskeletal hand model and a tendon-driven exoglove. A three-stage pipeline first extracts reference trajectories from video, then simulates hand impairments and object manipulation with a musculoskeletal model, and finally trains an RL-based exoglove controller to provide shared assistance. The results show the exoglove can compensate for muscle weakness, restoring roughly 90% of the healthy manipulation performance in simulation, with robust object-centric grasping across several YCB objects. This approach advances patient-specific adaptive control for upper-limb rehabilitation, while highlighting challenges in translating simulation-trained policies to real-world, vision-based settings.

Abstract

The use of wearable robots has been widely adopted in rehabilitation training for patients with hand motor impairments. However, the uniqueness of patients' muscle loss is often overlooked. Leveraging reinforcement learning and a biologically accurate musculoskeletal model in simulation, we propose a customized wearable robotic controller that is able to address specific muscle deficits and to provide compensation for hand-object manipulation tasks. Video data of a same subject performing human grasping tasks is used to train a manipulation model through learning from demonstration. This manipulation model is subsequently fine-tuned to perform object-specific interaction tasks. The muscle forces in the musculoskeletal manipulation model are then weakened to simulate neurological motor impairments, which are later compensated by the actuation of a virtual wearable robotics glove. Results shows that integrating the virtual wearable robotic glove provides shared assistance to support the hand manipulator with weakened muscle forces. The learned exoglove controller achieved an average of 90.5\% of the original manipulation proficiency.

Imitation Learning for Adaptive Control of a Virtual Soft Exoglove

TL;DR

The paper tackles personalized hand rehabilitation for motor-impaired users by combining imitation learning from healthy subject videos with a biologically accurate musculoskeletal hand model and a tendon-driven exoglove. A three-stage pipeline first extracts reference trajectories from video, then simulates hand impairments and object manipulation with a musculoskeletal model, and finally trains an RL-based exoglove controller to provide shared assistance. The results show the exoglove can compensate for muscle weakness, restoring roughly 90% of the healthy manipulation performance in simulation, with robust object-centric grasping across several YCB objects. This approach advances patient-specific adaptive control for upper-limb rehabilitation, while highlighting challenges in translating simulation-trained policies to real-world, vision-based settings.

Abstract

The use of wearable robots has been widely adopted in rehabilitation training for patients with hand motor impairments. However, the uniqueness of patients' muscle loss is often overlooked. Leveraging reinforcement learning and a biologically accurate musculoskeletal model in simulation, we propose a customized wearable robotic controller that is able to address specific muscle deficits and to provide compensation for hand-object manipulation tasks. Video data of a same subject performing human grasping tasks is used to train a manipulation model through learning from demonstration. This manipulation model is subsequently fine-tuned to perform object-specific interaction tasks. The muscle forces in the musculoskeletal manipulation model are then weakened to simulate neurological motor impairments, which are later compensated by the actuation of a virtual wearable robotics glove. Results shows that integrating the virtual wearable robotic glove provides shared assistance to support the hand manipulator with weakened muscle forces. The learned exoglove controller achieved an average of 90.5\% of the original manipulation proficiency.
Paper Structure (14 sections, 2 equations, 7 figures, 1 table)

This paper contains 14 sections, 2 equations, 7 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Learning manipulations from human videos and RL
  4. Model-based wearable robotics control
  5. Methodology
  6. Learning behaviour context from video data
  7. Musculoskeletal object manipulator
  8. Development of the exoglove controller
  9. Experiments
  10. Learning behaviour context from video data
  11. Training of hand-object manipulation tasks
  12. Learning of exoglove control
  13. Validation of exoglove biomechanics in simulation
  14. Discussion and Conclusion

Figures (7)

  • Figure 1: Pipeline overview illustrating the three main stages: (1) extraction of user behavior via imitation learning from healthy subject video data, (2) training a biologically accurate musculoskeletal hand model for object-centric manipulation and simulating muscle impairments, and (3) development and testing of a tendon-driven exo-glove controller to compensate for impaired hand functionality.
  • Figure 2: Visualization of the tendon-driven exoglove actuation in simulation, based on a real prototype rho2021learning. This figure visualizes the exoglove actuation mechanism, from left to right are the hand in normal state, finger contraction, and extension.
  • Figure 3: Visualization of the learned musculoskeletal model policy. The policy is initially trained as a motion prior, capturing general trajectory tracking capabilities, and is subsequently fine-tuned for specific downstream tasks. The figure illustrates the learned prior policy applied to manipulating a chef can (top row) and a sugar box (bottom row).
  • Figure 4: Prior model training with random initialization
  • Figure 5: Comparison of model training WITH (blue) and WITHOUT (red) the learned behaviour prior on previously unseen object and trajectory pair.
  • ...and 2 more figures