PaperlayerView exact revision on arXiv

Source-linked AI summary

Dressing in Motion: A Human Motion-Aware Diffusion Policy for Robot-Assisted Dressing

Haoxiang Sun, Fangyuan Wang, Songhao Huang, Justina Y. W. Liu, Jihong Zhu, Peng Zhou, David Navarro-Alarcon

arXiv:2609.04759v1cs.RO

TL;DR

Robotic dressing under human motion is difficult because garment–human contact and occlusion complicate action generation aligned with changing arm movements. The paper combines a diffusion policy learned from static demonstrations with PDE-based arm representation, region sampling, and registration for reactive trajectory adaptation. Across simulation and a real-world study, the method outperforms baselines in dressing progress, freedom of movement, and user comfort.

  • Problem

    Dynamic robotic dressing remains challenging because complex garment–human contact and occlusions make actions difficult to align with human arm movements.

  • Method

    A garment–human interaction-aware diffusion policy learns from partially observed point clouds, while PDE-based arm representation and point registration estimate motion-relevant regions for online trajectory adaptation.

  • Results

    The method outperforms baselines in dressing progress, freedom of movement, and user comfort across simulation and real-world evaluations.

  • Takeaways & Limitations

    The policy supports motion-aware dressing across nine participants, three garment types, and six arm-motion patterns.

Abstract

from arXiv · show

Robotic dressing assistance is a promising solution for supporting older adults with physical impairments in daily living. However, dressing under human motion remains challenging, as complex garment--human contact and occlusions make it difficult to generate actions aligned with arm movements. In this letter, we propose a visuomotor policy that learns dressing skills from static expert demonstrations and generalizes to dynamic user-motion scenarios. A diffusion policy tailored to garment--human interaction geometry learns from partially observed point clouds with varied arm postures. We then introduce an object-centric representation based on PDE diffusion to capture the axial distribution of the arm. By sampling motion-relevant regions and registering them across consecutive observations, the proposed method approximates arm motion and reactively adapts the executed trajectory. We evaluate our method in simulation and a real-world human study involving nine participants, three garment types, and six arm-motion patterns. Results show that our method outperforms baselines in dressing progress, freedom of movement, and user comfort. The project website is https://anonymous.4open.science/w/dressing-in-motion.

I. INTRODUCTION

Robotic dressing is motivated by caregiver shortages and the need to accommodate users’ natural arm movements, but deformable garments, occlusion, and close contact make dynamic dressing difficult. The paper proposes a diffusion policy with online arm-motion estimation and trajectory adaptation, evaluated across varied garments and motion patterns.

  • Dressing is physically intensive and repetitive, and caregiver shortages motivate robotic assistance for older adults and people with physical impairments.
  • Deformable garments, unpredictable occlusions, entanglement, tension, and possible gripper–body contact make dressing contact-rich and potentially uncomfortable or unsafe.
  • Most existing approaches assume a static arm, which simplifies occlusion handling but compromises user experience when users adjust posture.
  • The proposed policy combines diffusion-based action learning from static-arm demonstrations with online point registration that estimates arm motion and projects motion-compatible trajectories.
  • The method uses a diffusion policy tailored to garment–human interaction geometry and a PDE-diffused scalar field to localize motion-relevant arm regions.
  • Simulation and real-world evaluations cover diverse garment types and human motion patterns, with reported improvements in dressing progress, motion tolerance, and user ergonomics.

II. RELATED WORK

Prior robotic dressing research has used user modeling, visual or force feedback, model-based control, reinforcement learning, and imitation learning. The paper positions human-motion dressing as a difficult deformable-object interaction problem requiring adaptation to changing arm geometry and occluded observations.

  • Earlier work explored personalized user modeling, visual and force feedback, model-based dynamics, reinforcement learning, and imitation learning for robotic dressing.
  • Dressing is a contact-rich deformable-object manipulation task whose nonlinear material dynamics and infinite degrees of freedom complicate state estimation.
  • Learning from expert demonstrations is presented as a promising alternative because reinforcement learning still depends on faithful digital twins and reliable sim-to-real transfer.
  • Human-motion dressing requires adaptation to changing arm position and local contact-surface curvature, while improper contact can cause hand or elbow entanglement and occlusion creates cluttered observations.

III. PROBLEM FORMULATION AND ASSUMPTIONS

The task is formulated as partially observable visuomotor policy learning: the robot aligns and orients the sleeve before insertion, then advances it toward the shoulder while accommodating time-varying arm geometry after insertion.

  • The robot begins with the garment grasped in front of the arm, reduces hand–sleeve alignment error, and orients the sleeve for forearm insertion.
  • The objective is to advance the sleeve over the hand and elbow to the shoulder while accommodating time-varying arm geometry from partial point-cloud observations.
  • Before sleeve insertion the arm is assumed static, whereas after insertion the participant may vary the arm pose during garment advancement.

IV. METHOD

The method uses a hierarchical framework: a diffusion model learns general dressing actions from static demonstrations, while scalar-field-based point registration estimates arm transformations to align actions with changing configurations.

  • A diffusion model learns a general motion policy from static demonstrations spanning arm poses and garment types.
  • A diffused scalar field from occluded point clouds provides a continuous arm-axis coordinate for estimating arm-pose transformations through point registration.
  • The estimated transformations align generated motions with the current arm configuration and support reactive responses to non-Markovian human motion.

A. Diffusion Policy Learning from Demonstrations

The method learns a visuomotor diffusion policy from partially observed arm–garment point clouds and robot pose, using local interaction geometry to predict action chunks.

  • The policy fuses cropped arm–garment point-cloud features with encoded end-effector pose to form the observation conditioning action generation.The cropped points contain mixed human and garment points and are downsampled with farthest point sampling without additional segmentation.
  • The diffusion model iteratively denoises Gaussian noise conditioned on the fused observation to generate multimodal action distributions.The model predicts action chunks over a downstream execution horizon.
  • EdgeConv combines each point with local neighbor offsets to capture fine-grained human–garment interaction geometry.Max aggregation produces an N × d_h point-feature tensor before projection and fusion with the end-effector pose.
  • Training uses a standard noise-prediction objective for the diffusion policy.DDIM is used as the reverse-process sampling scheduler.

B. Arm-Centric Diffused Scalar Field

The method adapts dressing to changing arm configurations by extracting a motion-relevant arm region from occluded point clouds with PDE-based diffusion, then registering that region across observations.

  • The motion-adaptation problem is reduced to identifying the uncovered arm region and sampling the outermost points near the garment boundary along the arm axis.The selected region is intended to capture the surface governing sleeve–arm insertion behavior.
  • PDE diffusion assigns a smooth scalar field over observed points to encode geometry-aware axial structure for region sampling.The field is constructed from partially observed, cluttered point clouds rather than requiring reconstruction of the full occluded arm.
  • The full pipeline combines visual encoding and diffusion-policy action prediction with scalar-field region extraction and registration for reactive trajectory adaptation.The estimated transformation is applied to generate the adapted execution trajectory.
  • A shoulder source keypoint initializes diffusion so scalar values propagate along the arm surface.The source receives u(s, 0) = 1 while other observed points initially receive zero.
  • Short-time diffusion produces a scalar field sensitive to geodesic distance, allowing continuity and smoothness constraints to reject garment points and wrinkle artifacts.The resulting uncovered region U is used for subsequent region-of-interest sampling.

C. Arm-Motion-Aware Trajectory Adaptation

The controller estimates local arm motion by registering motion-relevant regions between observations and applies bounded corrections to pending end-effector targets.

  • The policy predicts an arm-centric end-effector trajectory chunk, then resamples the current region of interest at each execution step.The action chunk contains desired end-effector trajectories expressed in the arm-centric frame associated with the observation.
  • GICP registers the reference region to each current region to estimate the inter-frame transformation of the motion-relevant arm surface.The transformation includes rotation and translation estimated from point correspondences and local covariance information.
  • The estimated transformation provides a local rigid displacement used to project each pending position target through a bounded correction scheme.The correction adapts translation according to arm displacement while preserving the predicted orientation.

A. Numerical Simulation

The simulation evaluates dressing across varied arm motions, garment types, human models, and baselines. Results show that explicit trajectory adaptation and interaction-aware point-cloud encoding improve robustness as motion speed increases.

  • Experimental Setup: Four human models, three garment types, ten arm-motion patterns, and three speed profiles define the simulation evaluation.Arm speeds are static (v = 0), moderate (v = 1.0 rad/s), and fast (v = 2.0 rad/s).
  • Baselines: The comparison includes the full method, an adaptation-ablated variant, point-cloud and image-conditioned diffusion policies, Diff-MPC, and BC-LSTM.Ours w/o TA removes the human motion-aware trajectory adaptation module; DP3 tests interaction-aware point-cloud encoding.
  • Simulation Results: Under static conditions, point-cloud-conditioned diffusion policies outperform image-based policies by using richer 3D geometric cues.The reported comparison covers dressing performance in the simulation benchmarks.
  • Simulation Results: As arm-motion speed increases, most baselines degrade because they lack explicit motion adaptation, allowing hand motion to cause insertion failure or elbow entanglement.Motion can pull the arm away from the garment opening during alignment.
  • Simulation Results: Trajectory adaptation provides consistent gains across movement types, while interaction-aware point-cloud encoding improves generalization across garments and arm poses.These effects are identified through comparisons of Ours with Ours w/o TA and Ours w/o TA with DP3.
  • Simulation Results: The full method achieves average dressing ratio and sleeve insertion success rate above 0.95 and remains robust to human motion.The simulation results comprise 1470 trials, with Fig. 5 reporting dressing ratio and sleeve insertion success rate across motion speeds.

B. Hardware Experiments

The hardware study evaluates dressing across garments, participant motion patterns, subjective ratings, and motion-intensity conditions. The proposed method achieved strong real-world performance while showing degradation under the most intense arm motions.

  • Experimental setup: The study used a UR10e robot, overhead RealSense D435i observations, and vest, shirt, and windbreaker garments.The real-world evaluation involved independently collected and separately trained simulation and hardware datasets, without sim-to-real transfer.
  • Experimental setup: Nine participants completed trials involving varied garments and canonical arm motions, including up-and-down, side-to-side, waving, phone use, and object receiving.The canonical-motion data included 210 expert demonstrations collected from static subjects with varied arm configurations.
  • Subjective evaluation: Subjective evaluation measured freedom of movement and comfort on a six-level ordinal scale from 1.0 to 0.0, with higher scores indicating better performance.Freedom of movement captures arm restriction, while comfort primarily reflects perceived contact quality during dressing.
  • Results: The proposed method achieved a dressing ratio of 0.88 ± 0.105 and sleeve insertion success of 89% ± 10.4% across 252 real-world trials.The authors report that it outperformed all baselines in dressing progress, tolerance to human motion, and user-perceived comfort.
  • Results: Trajectory adaptation improved performance across movement types, while interaction-aware point-cloud encoding improved generalization across garments and arm poses.The comparison against Ours w/o TA supports the role of trajectory adaptation; comparison against DP3 supports the interaction-aware encoding benefit.
  • Motion-tolerance boundary: Performance degraded when arm-swing amplitude reached 80–100 cm and frequency reached 0.8–1.0 Hz.The capability test swept five amplitudes from 20–100 cm and five frequencies from 0.2–1.0 Hz across 150 trials.

VI. CONCLUSION

The paper presents a visuomotor policy for robot-assisted dressing under human arm motion. It combines diffusion-based action learning with arm-motion estimation and online trajectory adaptation, and reports generalization across novel arm configurations and user motions.

  • Conclusion: The proposed visuomotor policy accommodates human arm motion in robot-assisted dressing.It learns dressing actions from expert demonstrations conditioned on static arm configurations.
  • Conclusion: A garment–human interaction-aware diffusion policy uses a PDE-based arm representation, point-cloud sampling, registration, and online action projection for motion-aligned adaptation.The method evaluates simulation and real-world experiments involving nine participants, three garment types, and six arm-motion patterns.
  • Conclusion: The method generalizes to novel arm configurations and user motions while outperforming six baselines in dressing progress, sleeve insertion success, freedom of movement, and user comfort.These outcomes are reported across the paper’s simulation and real-world evaluations.
Loading 2609.04759v1…