Reproduce “FoldNet: Learning Generalizable Closed-Loop Policy for Garment Folding via Keypoint-Driven Asset and Demonstration Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuxing Chen, Bowen Xiao, He Wang
FoldNet introduces a framework for generating diverse synthetic garment data with semantic annotations and an iterative learning approach. This enables the training of robust closed-loop policies for garment folding, demonstrating direct sim-to-real transfer with a 25% increase in real-world success rates through improved error recovery.
Key insights
- Combining keypoint-controlled geometry with generative AI for texture synthesis and VLM-based filtering creates highly diverse and realistic synthetic garment assets efficiently.
- Explicitly training on automatically generated error recovery demonstrations, as with KG-DAgger, substantially enhances policy robustness and allows robots to recover from failures during real-world execution.
- High-quality synthetic data, especially when augmented with semantic annotations and error recovery examples, can effectively bridge the sim-to-real gap for complex deformable object manipulation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.