Ant Group's Robbyant open-sources LingBot-Touch: a tactile foundation model pretrained on ~5,000h of real touch data, Apache-2.0, lifting VLA success by 22–25 points

Why it matters

【Key points】

· 2026-11-09: Robbyant (Ant Group) open-sources LingBot-Touch, a JEPA tactile foundation model with one shared encoder for piezoresistive, magnetic and visuotactile sensors.

· Data: 4,983h of real touch from gloves and robot hands; 1.33B frames, 464K episodes.

· Weights: Tiny 5.46M and Small 21.76M params. Apache-2.0, ungated on HF.

· 26 real tasks on 2 hands: π0.5 48.1%→69.9%, LingBot-VLA 48.3%→73.1%, Diffusion Policy 21.6%→35.6%, for under 0.2% extra params.

· 12 sim RL tasks: 69.0% vs 21.4% without touch.

· Works on an unseen magnetic sensor (XHand1) with no retraining.

【Our read】

Background

· Touch data never pooled because every sensor differs. LingBot-Touch normalizes sensors before scaling data.

· Released: inference code and encoders. Not released: the corpus and training code.

Comparison

· vs Meta Sparsh / Sparsh-X (Digit 360): those are mostly single-modality; this is one encoder across three.

· vs FTP-1: claims over 10× lower contact error. Self-reported.

· vs NVIDIA GR00T 2 (e/1095): GR00T 2 is a world-action model, code due December. LingBot-Touch is a plug-in touch encoder you can download today.

Impact

· Raw touch gives at most half the gain and hurts in 22 of 104 pairs. The representation matters more than the sensor.

· Watch: independent replication, corpus release, formal paper.

【Verifiable】

• ·
• ·
• ·
• ·
• ·