Reproduce “Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map” on OpenResearch

About this paper

Lei Su, Zhijie Peng, Renyuan Ren, Shengping Mao, Juan Du, Kaifeng Zhang, Xuezhou Zhu

Tacmap introduces a framework for vision-based tactile sensing that unifies simulated and real-world tactile feedback through geometry-consistent penetration depth maps. This approach enables efficient, high-fidelity tactile simulation and successful zero-shot policy transfer for dexterous robotic manipulation tasks.

Key insights

• A shared, geometry-consistent penetration depth map serves as a robust and domain-invariant proxy for contact physics, effectively bridging the tactile sim-to-real gap.
• GPU-accelerated geometric ray-casting offers high-fidelity tactile simulation at a computational cost suitable for large-scale reinforcement learning, surpassing the efficiency of FEM methods.
• Aligning real-world tactile perception with simulation by supervising an image-to-map translation network with kinematically derived ground-truth depth maps (using the same geometric logic as simulation) is crucial for zero-shot transfer.

Summary by alphaXiv.