Reproduce “Dynamic Handover: Throw and Catch with Bimanual Hands” 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
Binghao Huang, Yuanpei Chen, Tianyu Wang, Yuzhe Qin, Yaodong Yang, Nikolay Atanasov, Xiaolong Wang
Researchers at UC San Diego and Peking University developed a Multi-Agent Reinforcement Learning framework that enables two multi-finger robot hands to dynamically throw and catch diverse objects. The system successfully transferred policies from simulation to real robots, demonstrating coordinated bimanual manipulation with a 0.60 success rate for ball objects in the real world.
Key insights
- Multi-Agent Reinforcement Learning (MAPPO) with partial observations promotes better generalization to novel objects compared to PPO with full state access in dynamic tasks.
- A firm initial dexterous grasp significantly improves the stability and repeatability of throwing trajectories, leading to more consistent object release.
- Integrating a real-time object trajectory prediction module (goal estimator) enhances robustness against physics disparities, allowing the catcher to adapt to unpredictable real-world dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.