Ant Group's Robbyant open-sources LingBot-Touch, a tactile foundation model for robot hands
The model was pretrained on 4,983 hours of magnetic, piezoresistive and visuotactile recordings and released with Apache-2.0 weights; the gains on 26 real-robot tasks are the company's own results, not yet independently replicated.
Why it matters
Most robot foundation models learn from vision and language, while touch data comes from many incompatible sensors. LingBot-Touch maps three sensor types into one shared Transformer through separate tokenizers, and Robbyant says it transfers to tactile hardware it never saw in pretraining. The company reports that, on two dexterous hands, all four tested action policies did best when fed LingBot-Touch tokens rather than raw tactile signals or no touch. Watch whether outside labs reproduce the gains and whether the separately distributed inference code appears.
Key facts
- Pretraining: 4,983 hours, 464,431 episodes, about 1.33 billion frames
- Sensors: magnetic, piezoresistive and visuotactile
- Smallest backbone: 5.46M parameters (HaT-Tiny)
- Tested on 26 real-world dexterous tasks, 20 trials each
- Weights released under Apache-2.0 on Hugging Face