Reproduce “Grasp to Act: Dexterous Grasping for Tool Use in Dynamic Settings” 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
Harsh Gupta, Mohammad Amin Mirzaee, Wenzhen Yuan
Researchers at the University of Illinois Urbana-Champaign developed the Grasp-to-Act (G2A) framework, integrating physics-based grasp optimization with an online reinforcement learning policy for continuous grasp adaptation, enabling dexterous robots to perform dynamic tool-use tasks. This approach achieved 100% grasp success and minimized translational and rotational slip across various tasks in both simulation and real-world experiments, outperforming multiple baselines.
Key insights
- Combining physics-based analytically stable grasp initialization with an online RL adaptation policy provides a more robust and efficient solution for dynamic manipulation tasks than either approach in isolation.
- Guiding grasp synthesis with human wrist pose information and rigorously evaluating stability under task-specific dynamic wrenches significantly improves the quality of initial grasps for tool use.
- Abstracting complex physical interactions into simplified force-based models, combined with extensive domain randomization, facilitates robust zero-shot sim-to-real transfer for dynamic, contact-rich tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.