Blind Dexterous Grasping via Real2Sim2Real Tactile Policy Learning

The article presents a novel framework for tactile-only blind grasping using a multi-fingered robotic hand, addressing challenges in sim-to-real transfer. Key innovations include a Real2Sim tactile calibration pipeline for accurate tactile signal reproduction, a layout-aware tactile encoder for enhanced observation expressiveness, and a tactile-conditioned Diffusion Policy trained on successful grasp trajectories. The resulting system demonstrates a 27% success rate in real-world grasping across various objects, highlighting the effectiveness of combining calibration, geometry-aware learning, and policy aggregation for improving tactile manipulation in robotics.