Reproduce “Grasping in the Wild:Learning 6DoF Closed-Loop Grasping from Low-Cost Demonstrations” 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
Shuran Song, Andy Zeng, Johnny Lee, Thomas Funkhouser
Researchers from Columbia University and Google developed a low-cost hardware system for collecting human grasping demonstrations and a 6DoF closed-loop vision-based grasping algorithm leveraging this data. Their approach achieves up to 92% grasping success on novel objects in static scenes and 88% in dynamic settings, while significantly improving learning efficiency.
Key insights
- Diverse human grasping demonstrations collected with a low-cost, robot-mimicking device can effectively bootstrap complex 6DoF robot learning, overcoming data scarcity and improving generalization.
- An "action-view" based rendering mechanism, integrated with a continuously updated TSDF, efficiently models visual outcomes of actions, simplifying the learning of value functions for 6DoF closed-loop control.
- Combining human demonstration pretraining with real robot fine-tuning significantly accelerates learning convergence and achieves higher final performance compared to learning from scratch, effectively bridging the simulation-to-reality gap.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.