$\mathcal{N}_0$-Foundation: Towards the Age of Tactile Intelligence
The paper introduces a tactile manipulation stack built around 30,000 hours of paired visual and touch data.
The authors describe new hardware and collection infrastructure, including a vision-based tactile sensor, a tactile UMI, and synchronized visuo-tactile recording for robots and demonstrations. Their NeoData corpus spans six embodiments, 450 tasks, and billions of paired RGB and tactile frames, with a 5,000-hour open subset called OpenNeoData. They also propose NeoForce, a representation model meant to transfer tactile signals across sensor designs, plus NeoReal and NeoSim benchmarks. Experiments reported in the abstract say policies gain from physical contact state rather than device-specific tactile appearance. ArXiv · AI/CL/LG's note
The authors describe new hardware and collection infrastructure, including a vision-based tactile sensor, a tactile UMI, and synchronized visuo-tactile recording for robots and demonstrations. Their NeoData corpus spans six embodiments, 450 tasks, and billions of paired RGB and tactile frames, with a 5,000-hour open subset called OpenNeoData. They also propose NeoForce, a representation model meant to transfer tactile signals across sensor designs, plus NeoReal and NeoSim benchmarks. Experiments reported in the abstract say policies gain from physical contact state rather than device-specific tactile appearance. ArXiv · AI/CL/LG's note
score 6