Reproduce “TossingBot: Learning to Throw Arbitrary Objects with Residual Physics” 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
Andy Zeng, Shuran Song, Johnny Lee, Alberto Rodriguez, Thomas Funkhouser
A robotic system learns to grasp and throw diverse rigid objects into target bins, leveraging a "Residual Physics" approach that blends analytical ballistic models with learned neural network corrections. The system demonstrates a high throughput of 514 picks per hour and 84.7% throwing accuracy, while generalizing to novel objects and target locations.
Key insights
- Throwing significantly boosts robotic pick-and-place efficiency, achieving higher throughput and extending the functional workspace beyond the robot's direct kinematic reach.
- Jointly optimizing grasps based on downstream throwing accuracy enables the robot to learn stable grasps that are more conducive to successful and predictable throws.
- The "Residual Physics" hybrid approach effectively combines the generalizability of physics models with data-driven corrections for unmodeled dynamics, leading to robust performance and strong generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.