Reproduce “Learning In-Hand Translation Using Tactile Skin With Shear and Normal Force Sensing” 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
Jessica Yin, Haozhi Qi, Jitendra Malik, James Pikul, Mark Yim, Tess Hellebrekers
Researchers from Meta AI, UPenn, UC Berkeley, and UW-Madison developed an efficient, 3-axis tactile skin model for the ReSkin sensor, enabling zero-shot sim-to-real transfer of binary normal and ternary shear force signals. This model facilitates training reinforcement learning policies for dexterous in-hand object translation, demonstrating significant improvements in task performance and adaptation to diverse out-of-domain objects and environmental conditions.
Key insights
- Developing an RL-tractable tactile skin model that balances simulation speed and fidelity for both normal and shear forces is crucial for bridging the sim-to-real gap in dexterous manipulation.
- Rich 3-axis tactile feedback, particularly signed shear forces, enables policies to form more stable internal representations of object state and interaction dynamics, leading to enhanced robustness and generalization to diverse out-of-domain conditions.
- Tactile sensing empowers robots to explore a wider range of adaptive finger gaits and motor control strategies, essential for complex, contact-rich tasks like in-hand object translation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.