Reproduce “Right-Side-Out: Learning Zero-Shot Sim-to-Real Garment Reversal” on OpenResearch

About this paper

Chang Yu, Siyu Ma, Wenxin Du, Zeshun Zong, Han Xue, Wendi Chen, Cewu Lu, Yin Yang and 4 more

Researchers at UCLA and Toyota Research Institute introduced a zero-shot sim-to-real framework for the underexplored task of autonomously reversing inside-out garments to right-side-out using robotic manipulators. The system, which decomposes the task into sequential sub-goals and uses a custom GPU-parallel Material Point Method (MPM) simulator for training, achieved an 81.3% success rate on diverse real-world garments without any real-world training data.

Key insights

• Decomposing a complex, long-horizon deformable object manipulation task into simpler, interpretable primitives significantly enhances robustness and learning efficiency, outperforming end-to-end approaches for contact-rich scenarios.
• High-fidelity simulation, particularly with an efficient MPM solver, is critical for accurately modeling dynamic, self-colliding cloth interactions, enabling scalable data generation and robust zero-shot sim-to-real transfer.
• Real-world perception noise is a major barrier to zero-shot transfer; applying a depth refinement module like D3RoMa is essential for stabilizing keypoint predictions and ensuring reliable robotic grasp and manipulation actions.

Summary by alphaXiv.