Robbyant's LingBot-Touch cuts cross-sensor tactile error by ~97% on unseen robot hand
omarsar0 · x · 2026-11-11
Robbyant (Ant Group) released LingBot-Touch, a tactile foundation model tackling a key robotics pain point: models trained on one tactile sensor rarely transfer to another.
- Pretrained on 4,983 hours of real touch data, mapping every sensor type (piezoresistive, magnetic, visuotactile) to a unified virtual fingertip representation, with varied touch-point density and placement to prevent memorizing a single sensor.
- On the unseen XHand1 hand, the frozen model plus a small add-on for fingertip localization achieved 0.16 mm average contact-center error vs 5.89 mm for FTP-1 with a hand-specific adapter — a 97% reduction.
- Tactile tokens are plug-and-play with diffusion policy and VLA models (π₀, LingBot-VLA), improving dexterous manipulation across real robot deployments. Code and checkpoints are open-sourced.
More from Embodied
- World's first fighting teleop: S.T.E.E.L. brings VR and mocap-driven real-time sparring — cixliv · 2026-11-11
- EEG can't match implanted BCIs: physics caps scalp signals at ~70Hz — EricMTrautmann · 2026-11-11
- Dyna Robotics' CoRL 2026 demo blends large-scale navigation with dexterous manipulation — JasonMa2020 · 2026-11-11
- MBZUAI releases PAN 2, an open-weights foundation world model for physical intelligence — Scobleizer · 2026-11-11
- Hobbyist builds MCP server into BIOS programmer so Claude Code can debug corrupted firmware — AsianKidTV · 2026-11-11
- SpaceX locks 100+ GW of solar cell equipment from SINGULUS for Starmind AI satellites — ns123abc · 2026-11-11