Force-Constrained Visual Policy: Safe Robot-Assisted Dressing via Multi-Modal Sensing
This research paper introduces the Force-Constrained Visual Policy (FCVP) technique for robot-assisted dressing, combining a simulation-trained vision policy with a real-world-learned force dynamics model. By framing the task as a constrained optimization problem, the system ensures that garments are successfully maneuvered onto a human arm while filtering out actions that exceed a safe force threshold. Evaluations through both simulation and a 240-trial real-world human study with 10 participants demonstrate that FCVP significantly reduces force violations compared to existing baselines while maintaining high dressing performance.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
Robotics researchers and AI developers working on safe human-robot interaction and assistive technologies should read this to understand multi-modal sensorimotor learning.
What the document contains
The document addresses the development and evaluation of the Force-Constrained Visual Policy (FCVP) method for single-arm robot-assisted dressing. It details how a vision-based policy trained via reinforcement learning in simulation can be paired with a force dynamics model learned directly from real-world data to filter out unsafe actions.
The paper does not provide legislative regulations, commercial deployment compliance standards, or legal frameworks for assistive medical devices.
- Robot-assisted dressing benefits from combining both visual sensing for garment and body pose observation and force sensing for safety and comfort.
- Simulation provides sufficient accuracy for point cloud data and vision policies but fails to accurately simulate force dynamics for deformable garments interacting with humans.
- The proposed Force-Constrained Visual Policy (FCVP) successfully combines a simulation-trained vision policy with a real-world-trained force dynamics model using constrained optimization.
- Real-world human studies with 10 participants across 240 trials show that FCVP significantly reduces force violations compared to vision-only or unconstrained multimodal approaches.
· I. Introduction · II. Related Work · III. Problem Statement and Assumptions · IV. Background - Vision-based Policy Learning in Simulation · V. Method · VI. Experiments · VII. Conclusion
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Accountability
0/5Clear responsibility for AI outcomes
N/A
screening score · 2026-10-24
Fairness
0/5Non-discrimination and equitable treatment
N/A
screening score · 2026-10-24
Human Dignity
1/5Respect for the inherent worth of persons
Robot-assisted dressing could profoundly enhance the quality of life of adults with physical disabilities.
Abstract
screening score · 2026-10-24
Human Oversight
4/5Humans remain in control of consequential decisions
We run each trial for a fixed number of time steps, unless the participant asks to stop or the measured force is above a safety threshold (15 Newtons).
Section VI-B
screening score · 2026-10-24
Privacy
0/5Protection of personal data and private life
N/A
screening score · 2026-10-24
Transparency
1/5Openness about how AI systems work and are used
Video demonstrations are available on our project website
Abstract
screening score · 2026-10-24
Enforcement
0/5Consequences and compliance mechanisms with real force
N/A
screening score · 2026-10-24
Impact Assessment
3/5Required assessment of an AI system's risks/impacts before or during use
Undesired motions performed by the robot that strain the garment or cause accidental collisions with a person could apply large forces to the human body and pose discomfort and potential safety risks.
Section I
screening score · 2026-10-24