Deform360 is a large-scale real-world visuotactile dataset for modeling deformable objects in robotic manipulation. It contains 198 everyday objects, 1,980 interaction sequences, and more than 215 hours of observations captured with 41 surround-view cameras and bimanual tactile grippers. The authors introduce a markerless visuotactile 3D tracking pipeline that extracts dense geometry and motion, then benchmark 2D video-based world models against 3D particle-based models. The paper also presents a preliminary robot-planning demonstration, highlighting trade-offs between structural priors and scalability.
No heat snapshots are available in the last 24 hours.