Reproduce “ConCent: Contact-Centric Real-to-Sim-to-Real Learning from One Demonstration” 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
Heecheol Kim, Namiko Saito, Katsushi Ikeuchi, Yasuyuki Matsushita
ConCent is a real-to-sim-to-real reinforcement learning framework designed for contact-rich robot manipulation that utilizes a single real-world demonstration. It achieves an 80.0% success rate on a challenging shape sorter insertion task by explicitly aligning simulated local contact dynamics with real observations and enforcing demonstrated contact event sequences, alongside a Virtual Collision Penalty (VCP) mechanism that enhances training scalability by 1.3x.
Key insights
- Robust sim-to-real transfer for contact-rich tasks relies on accurately reproducing task-relevant local contact dynamics and their temporal sequences, rather than achieving perfect global physics fidelity.
- A single real-world demonstration can provide sufficient structured physical information to ground simulation and automatically derive a precise, physically plausible reward for RL.
- Computational efficiency in contact-rich simulations can be drastically improved by prioritizing and fully simulating only essential contact points, while penalizing proximity for non-essential ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.