Reproduce “One Policy to Dress Them All: Learning to Dress People with Diverse Poses and Garments” on OpenResearch

About this paper

Yufei Wang, Zhanyi Sun, Zackory Erickson, David Held

Researchers from Carnegie Mellon University developed a robot-assisted dressing system capable of generalizing to diverse garments and human arm poses using only partial point cloud observations. The system achieved an average whole arm dressed ratio of 0.86 1 0.17 and an upper arm dressed ratio of 0.71 1 0.34 in a human study with 17 participants, demonstrating enhanced performance over a non-distilled baseline.

Key insights

• A segmentation-type PointNet++ architecture (Dense Transformation policy) is effective for learning dense actions from partial point clouds for complex deformable object manipulation tasks like dressing.
• Policy distillation using Earth Mover's distance provides a robust method for learning a single generalizable policy across a wide range of diverse conditions (e.g., human poses) by leveraging specialized teacher policies.
• Guided domain randomization successfully bridges the sim2real gap by training a student policy with randomized inputs while using non-randomized teacher policies for imitation, yielding policies robust to real-world sensor noise and discrepancies.

Summary by alphaXiv.