Robbyant open-sources LingBot-Touch tactile model for robots

Robbyant's release addresses a practical bottleneck in robot touch: commercial tactile sensors vary widely in how they capture contact, and a model trained on one sensor type often struggles to reuse that experience when the hardware changes. LingBot-Touch tackles this by providing separate encoding pathways for magnetic, visual and piezoresistive sensing, then mapping contact points into a shared fingertip coordinate space. That design lets data from human-worn tactile gloves and teleoperated hands train the same model, and allows different action models to extract contact features without being tied to Robbyant's own policies.

The performance figures come from 26 real-world tests across two dexterous hands, Sharpa Wave and XHand. On XHand, the magnetic tactile sensor was not part of pre-training; all magnetic experience came from human glove data, and the team kept LingBot-Touch parameters unchanged while training only the action model — a test of whether the learned representation transfers across hardware. The lightweight Tiny version, at about 5.3 million parameters, was used for all 26 tests, suggesting the approach is intended to fit on-device compute constraints rather than only in simulation.

By open-sourcing the model and technical report, Robbyant positions LingBot-Touch as a general-purpose perception layer between sensors and action models, rather than a module for a specific platform. That could lower integration costs for hardware vendors and robotics teams, though the body does not detail commercial terms or availability beyond Hugging Face and GitHub.