Ant's LingBot-Touch: open-source tactile foundation model boosts dexterous manipulation across 26 tasks
新智元 · wechat · 2026-11-10
Ant Group's Lingbo (Robbyant) open-sourced LingBot-Touch, a universal tactile foundation model that unifies vision-based, magnetic and piezoresistive tactile signals into generic "tactile tokens" usable by any action model.- Lightweight plug-in: the Tiny version has just 5.32M params, adding only 0.13-0.16% to three major action models; spray-bottle grasping rose from 42.5% to 87.5%, blind Mahjong sorting hit 70%.- Data & method: 5,000 hours / 1.33B frames of real tactile data, two-thirds from humans wearing tactile gloves; magnetic-tactile data transferred zero-shot to a new dexterous hand, lifting π0.5 from 48% to 74.17%.- Architecture: per-sensor encoders into a shared trunk, 5D-RoPE fingertip mapping, and Tactile-JEPA that masks 90% null readings and predicts forces instead, lifting retrieval recall to 49.32%.- Results: across 26 real-robot and 12 sim tasks, T-Rex, π0.5 and LingBot-VLA gained +23.18/+21.81/+24.81 points on average; blind tactile localization error down to 0.12mm. Scaling trends were observed as data grew to 5,000 hours. The article contrasts this with GPT-6 Astra breaking real hardware (0/33 on plug tasks) due to lacking touch feedback.
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