15-Institution TouchScale Dataset: 500 Hours of Vision-Tactile Data Doubles Robot Success Rate
机器之心 · wechat · 2026-10-09
Researchers from Texas A&M, Google DeepMind, CMU, Stanford, NVIDIA, Meta and 10 other institutions released TouchScale: a 500-hour human vision-tactile dataset collected with a single unified wearable system (880 taxels per glove, 87k interactions, 1,500+ objects), solving the confounding issue of merging heterogeneous datasets.
Key results:
- Human 'feel' transfers to robots: using human visual-tactile data only for mid-training (no action labels, no retargeting), an xArm6 + BrainCo dexterous hand improved average success on 4 contact-rich tasks from 22.5% to 57.5% (e.g., soft/hard sorting 10%→60%, bottle-cap removal 40%→70%).
- Scaling gains emerge: growing data from 10% to 100% lifted zero-shot cross-sensor contact prediction cIoU from 0.311 to 0.383; at equal 16-hour training budget TouchScale beats EgoTouch (0.181 vs 0.134), and tactile supervision also improves visual representations.
Paper: arxiv.org/abs/2610.10288; dataset open-sourced on Hugging Face. The authors argue tactile learning is now starting its own scaling journey, mirroring what vision went through.
More from Embodied
- Atomic Machines exits stealth with AI-native Matter Compiler that builds micromachines from code — DeryaTR_ · 2026-10-10
- Datawhale's Embodied AI Tutorial Hits 4k Stars: Build VLA from Scratch to Real-Robot Deployment — philfung · 2026-10-10
- Carmack tries Waymo and Zoox: sparse pickup points are the biggest pain point — ID_AA_Carmack · 2026-10-10
- Archon Robotics shows off Whole-Body Intelligence framework pushing humanoids toward real-world deployment — jiqizhixin · 2026-10-10
- Six-month-old Archon Robotics demos whole-body intelligence: humanoid fetches toys by climbing a step stool — 机器之心 · 2026-10-10
- ROMA: a 7B model that lets robots understand objects by shaking, weighing and touching them — 机器之心 · 2026-10-10