Robbyant's LingBot-Touch: a tactile foundation model trained on 5,000 hours of real data
机器之心 · wechat · 2026-11-10
Robbyant's LingBot-Touch is a tactile foundation model trained on 4,983 hours of real contact data (1.33B frames) spanning vision-based, magnetic, and piezoresistive sensors, using multi-branch tokenizers over a shared Haptic Transformer with a Tactile-JEPA latent-prediction objective and a unified 3D fingertip coordinate system. Plugged into four action models (Diffusion Policy, π0.5, LingBot-VLA, T-Rex) across 26 real-world dexterous-hand tasks, it lifts success rates substantially (e.g. π0.5 from 47.56% to 68.58% on SharpaWave); scaling tests show gains plateau at 5,000 hours unless model capacity grows. The article also notes Anthropic's Claude Fable 5.2 driving robots to solve Rubik's cubes and score 64.9% on RoboDojo.
More from Embodied
- Archon Robotics unveils Anchor 1, a whole-body humanoid controller with lowest tracking error — Scobleizer · 2026-11-10
- LingBot-Touch: a foundation model for heterogeneous tactile representation learning lands on HF — _akhaliq · 2026-11-10
- Alex Kendall's Team Releases Three-Part Series on Safety and Validation in End-to-End Learning — alexgkendall · 2026-11-10
- Stanford BEHAVIOR Challenge: robot full-task success jumps from 12% to 48% in one year — yining_hong · 2026-11-10
- Amazon FAR hiring researchers and PhD interns in SF for human-level open-world navigation — iamfakhrealam · 2026-11-10
- Apple is building a world-class robotics team, recruiting at CoRL — iamfakhrealam · 2026-11-10