Reproduce “Learning to Throw-Flip” 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
Yang Liu, Bruno Da Costa, Aude Billard
Researchers at EPFL developed a robotic system capable of "throw-flipping" objects to achieve precise landing positions and orientations, reducing sample complexity by 40% through physics-informed learning and accelerating learning for new objects by 70% via transfer learning.
Key insights
- The impulse-momentum-based control design and the 'damping' parameter effectively expand the robot's reachable landing pose space, enabling full 360-degree object rotations.
- Integrating known projectile dynamics into the learning process via a projectile-based forward model significantly reduces the sample complexity required to learn new throwing tasks.
- Transfer learning, by leveraging existing knowledge about in-hand spinning behavior, dramatically accelerates the acquisition of throwing skills for objects with shifted Centers of Mass.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.