Ant Group's Robbyant releases tactile model LingBot-Touch

Ant Group's robotics division, Robbyant, has released the tactile model LingBot-Touch under the Apache-2.0 license. The model was trained on 4,983 hours of recordings from three types of sensors and validated on 26 real-world bimanual tasks.

Compiled by Sergey KostenchukPublished 2026-11-11Updated 2026-11-11

2026-11-11 Research HuggingFace

🤖 Ant Group's Robbyant releases tactile model LingBot-Touch

Ant Group's division has released the weights and inference code for the tactile model LingBot-Touch (Apache-2.0) along with a technical report arXiv:2611.12806. The JEPA model predicts latent representations of tactile signals, was trained on 4,983 hours of recordings from three types of sensors, and was validated on 26 real-world bimanual tasks.

🌍 Typically, a separate model is trained for each type of tactile sensor. LingBot-Touch maps piezoresistive, magnetic, and visuotactile signals into a common latent space—checkpoints can be used as a base. Transfer is currently confirmed only for Sharpa Wave and XHand.

👤 The Robbyant/lingbot-touch repository contains inference code, examples for Sharpa and XHand, and instructions (Python 3.12, PyTorch 2.11, CPU). Tiny and Small checkpoints are available on Hugging Face and ModelScope.

Source 1: https://github.com/Robbyant/lingbot-touch Source 2: https://technology.robbyant.com/lingbot-touch

Sources