Reproduce “COCOI: Contact-aware Online Context Inference for Generalizable Non-planar Pushing” 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
Zhuo Xu, Wenhao Yu, Alexander Herzog, Wenlong Lu, Chuyuan Fu, Masayoshi Tomizuka, Yunfei Bai, C. Karen Liu and 1 more
A deep reinforcement learning framework, COCOI, enables robots to infer contact dynamics context online from multi-modal sensors for generalizable non-planar pushing tasks. The system demonstrated a 90% success rate in real-world deployments by adapting to diverse objects and dynamics, including unseen ones.
Key insights
- Online inference of a latent dynamics context representation through contact-rich interaction with multi-modal sensors significantly improves policy generalization across diverse object properties.
- Actively sampling interaction history when contact forces are present is more effective for learning relevant dynamics context than uniform sampling.
- The learned context representations effectively encode structured dynamics information, allowing the policy to adapt to unobservable object properties like mass distribution and friction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.