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
cs.RO
2026-09-04
A diffusion policy trained on static demos adapts online to arm motion. Nine people, three garments, six motions: dressing ratio 0.88 and 89% insertion over 252 trials.
Most robot-assisted dressing systems assume the arm stays still. A static arm is a gift under occlusion: hand, elbow, and shoulder can be inferred from the start pose. Once the person moves, the sleeve misses the hand, fabric wraps the elbow, and the gripper can hit the body. Dressing is also contact-rich deformable manipulation, so a compact state is hard to write down. This group at The Hong Kong Polytechnic University trains on static-arm demonstrations and executes while the arm is allowed to move.
Two layers.
A diffusion policy learns dressing actions from static expert demos. The observation is a point cloud cropped around the dressing-side arm, downsampled by farthest-point sampling, with no extra segmentation: human and garment points stay mixed. One EdgeConv layer encodes each point plus neighbor offsets, capturing local garment-arm geometry, then fuses with the end-effector pose. DDIM denoises an action chunk. Prediction runs at Tp=8 and 15 Hz; execution uses Ta=4 and needs above 50 Hz, so the policy emits chunks slowly and a second loop adapts them fast.
The lower loop does not rebuild the full arm. Dressing is assumed to be governed by the still-uncovered arm points closest to the current sleeve boundary along the arm axis. A short-time PDE diffusion is solved on the cloud, with the shoulder as a source, producing a smooth scalar field along the surface. That field drops garment wrinkles, isolates the uncovered region, and samples a region of interest R. Consecutive R clouds are registered with GICP to get a local rigid transform. Only the translational target is warped; orientation is kept. A bounded scale on the transform stops violent corrections.
The arm is required to stay still before sleeve insertion and may move after. Force sensing is out of scope and listed as future work.
Simulation is Assistive Gym: four body shapes, three garments cropped from Cloth3D, ten arm-motion patterns, three speeds (static / 1.0 / 2.0 rad/s). Imitation methods share 180 teleoperated trajectories; Diff-MPC uses 750 in its original setup. Over 1470 trials, the full method stays above 0.95 on both dressing ratio and sleeve-insertion success. Baselines drop once the arm speeds up, because they have no online correction.
Hardware is a UR10e with a RealSense D435i. Garments are a vest, a shirt, and a windbreaker. 210 expert demos are all static, with varied poses. Sim and real use separate datasets and separately trained policies; there is no sim-to-real transfer. Nine participants (three female, six male, 162-185 cm) each run 28 trials, 36 per method, 252 in total. Six daily motions: up-down, side-to-side, back-and-forth, waving, phone use, receiving an object. Safety caps are 20 kg·m/s joint momentum and 120 N tool force.
| Method | Vest | Shirt | Windbreaker |
| Ours | 0.92 | 0.89 | 0.83 |
| w/o trajectory adaptation | 0.93 | 0.92 | 0.50 |
| DP3 | 0.79 | 0.58 | 0.67 |
| DP-image | 0.45 | 0.33 | 0.25 |
| Diff-MPC | 0.14 | 0.30 | 0.27 |
| BC-LSTM | 0.58 | 0.50 | 0.46 |
Across 252 trials: dressing ratio 0.88±0.105, success 89%±10.4%. Four planar motions sit at 1.00; phone use falls to 0.77; receiving an object is 0.50. Removing trajectory adaptation drops the windbreaker from 0.83 to 0.50, so the correction mainly helps long sleeves under motion. A 150-trial sweep of amplitude and frequency starts to fail at 80-100 cm and 0.8-1.0 Hz. On a laptop RTX 5070, DDIM averages under 50 ms and 96-point GICP under 10 ms.
Static demos and dynamic execution are split: the policy does not need motion demonstrations, and online registration projects the planned chunk onto the current arm. For assistive dressing, the user cannot be asked to freeze. DP3 is the closest point-cloud diffusion baseline and still lags on motion. The numbers come from a human study, not simulation alone.
The hard cases are honest. A windbreaker and an object-receiving motion are not solved. Force feedback, bimanual dressing, and recumbent users are outside the setup.
The authors flag two limits. Motion before insertion can ruin hand-sleeve alignment, partly because control rate cannot match human motion. Mapping arm motion to end-effector translation injects irregular displacement; the bounded correction trades jitter against tracking lag, and large fast swings escape it.
Force is unused; contact safety is a momentum and force threshold. The scalar field is sourced at the shoulder, with no separate stress test for a blocked shoulder or a badly broken cloud. Participants are healthy adults sitting down, 162-185 cm, not older adults or people with motor impairment who would actually need the system. Thirty-six trials per method, split across garments and motions, leave the 0.50 object-receiving cell thin.