Ant Lingbo open-sources LingBot-Touch: tactile tokens lift robot zipper success from 15% to 59%

赛博禅心 · wechat · 2026-11-10

Ant Group's Lingbo team open-sourced LingBot-Touch, a tactile model for robots (code and weights on GitHub/HuggingFace). It unifies readings from vision-based, magnetic, and piezoresistive tactile sensors into a common "tactile token" representation for action models — fixing the problem where raw sensor data actually hurts performance (SharpaWave+π0.5 zipping dropped from 39% to 15% with raw readings).

Key facts:

Trained on 4,983 hours of real-world tactile data (2/3 human-worn gloves, 1/3 teleoperated hands), using Tactile-JEPA: masking some sensor points and predicting masked contacts from the rest.
Across 26 real-robot tasks and 4 action models: bottle-cap turning reached 90%; Mahjong-tile sorting (identifying tiles by touch only) rose from 5% to 56.25%; SD-card ejection on T-Rex from 32% to 92%.
Cross-hand transfer: works on the out-of-training XHand from Robot Era — π0.5's average success rose from 48% to 74.17%, so vendors can pick any sensor and reuse this tactile layer.
The model is only 0.00532B parameters for low latency (a larger Small variant available); the team next plans to fuse vision, depth, touch, and temperature into one model and extend touch from fingertips to the whole hand.

Original post →

Ant Lingbo open-sources LingBot-Touch: tactile tokens lift robot zipper success from 15% to 59%

More from Embodied