Reproduce “CoRMA: Contrastive RMA for Contact-Rich Meta-Adaptation” 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
Wentian Wang, Chutong Wen, Hongxu Ma, Wuhao Wang, Zhexiong Xue, Abdul Haseeb Nizamani, Dandi Zhou, Xinhai Sun and 1 more
CoRMA introduces a meta-adaptation framework that uses a causal Transformer and a force-regime contrastive objective to predict a semantic contact context from deployable sensor history. This approach significantly reduced the sim-to-real performance degradation gap by up to 76.25% (e.g., from -86.25% to -10% for PegInsert) compared to a strong baseline on contact-rich robotic assembly tasks like peg insertion, gear meshing, and nut threading.
Key insights
- A compact, semantically meaningful representation of contact states (`Privileged Z`) derived from simulator dynamics provides effective and reusable supervision for adapting to contact-rich tasks.
- Leveraging force, proprioception, and action histories directly as adaptation evidence is more effective for interpreting physical interaction in contact-rich scenarios than relying on visual or object-centric cues.
- Integrating a force-regime contrastive objective regularizes the adapter's representation, allowing it to group similar contact meanings across different tasks and enhance interpretation of ambiguous force signals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.