LingBot-Touch
Ant Group's Robbyant open-sourced LingBot-Touch, a tactile foundation model pretrained on nearly 5,000 hours of data from three sensor types, for robot hands.
LingBot-Touch gives each tactile sensor type (magnetic, visuotactile, piezoresistive) its own input tokenizer feeding a shared Transformer, and maps contacts into a common fingertip space so touch data can transfer across hardware. Robbyant reports that adding its representation raised average task success by about 14 to 26 percentage points for four action models (LingBot-VLA, π0.5, Diffusion Policy, T-Rex) on 26 real-robot tasks. The release is inference only, with the frozen encoder and feature extraction utilities.
- Date
- Tuesday 10 November 2026
- Lab
- Robbyant (Ant Group)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Pretraining data | nearly 5,000 hours real tactile recordings from human gloves and teleoperated hands | company |
| Average task success gain across four action models | 14 to 26 percentage points 26 real-robot tasks on Sharpa Wave and XHand, versus no touch | company |
| T-Rex cap twisting on Sharpa Wave | 90% versus 56.25% without touch and 62.5% with raw sensor signals | company |
| Tiny model size | about 5 million parameters adds about 0.13% to 0.16% to the policies | company |
Code under Apache-2.0 with Tiny and Small checkpoints on Hugging Face and ModelScope; all results are Robbyant's own with no independent replication.
Sources
- pandaily.com/ant-group-robbyant-lingbot-touch-open-source-tactile-foundation-model
- technode.com/2026/11/10/ant-groups-robbyant-open-sources-tactile-ai-model-for-robot-hands/
This record was checked against its sources on 10 November 2026. How we check