Reproduce “In-Hand Object Rotation via Rapid Motor Adaptation” 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
Haozhi Qi, Ashish Kumar, Roberto Calandra, Yi Ma, Jitendra Malik
A rapid motor adaptation framework enables a multi-fingered robot hand to perform generalized in-hand object rotation around the z-axis. The system, trained in simulation using only cylindrical objects, directly transfers to a real robot and robustly rotates over 30 diverse, unseen objects by estimating object properties solely from proprioceptive feedback.
Key insights
- The Rapid Motor Adaptation paradigm, utilizing low-dimensional "extrinsics" to encode environmental (object) properties, is effective for online adaptation in complex, contact-rich in-hand manipulation tasks, extending its utility beyond locomotion.
- Proprioceptive sensing alone (robot joint positions and motor commands) is sufficient for estimating crucial object properties and enabling robust generalized in-hand rotation, significantly reducing sensing requirements compared to systems relying on vision or tactile feedback.
- Natural, stable, and high-clearance finger gaits for object rotation can emerge automatically from reinforcement learning with a well-designed composite reward function, without requiring explicit behavioral heuristics or predefined contact sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.