Robbyant (Ant Group) · Tactile foundation model

LingBot-Touch

LingBot-Touch is a JEPA-based tactile representation model from Robbyant. It uses a shared Haptic Transformer with modality-specific tokenizers to learn unified representations from piezoresistive, magnetic and visuotactile sensors. It was pretrained on 4,983 hours of recordings (about 1.33 billion frames across 464K episodes), and the smallest encoder, HaT-Tiny, has 5.46M parameters. The page reports gains when frozen LingBot-Touch features are added to policies on 26 real-world dexterous tasks (Sharpa Wave and XHand hands, 20 trials per task) and 12 simulated reinforcement learning tasks, with performance improving as pretraining data grows from 0 to 5K hours. A tech report, code, and Hugging Face and ModelScope links are listed; the page gives no licence.

Open the source ↗

Similar Agents & Automation releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.