Robbyant open-sources LingBot-Touch tactile model for dexterous robots
Robbyant says its open LingBot-Touch model helps robots transfer tactile experience across sensors, platforms and dexterous manipulation tasks.
Robbyant has open-sourced LingBot-Touch, a tactile representation model designed to help dexterous robots transfer what they learn across sensor types, hardware platforms and manipulation tasks. The project describes it as a JEPA-based model built around a shared Haptic Transformer, with separate tokenisers feeding a common latent representation of contact and motion 1.
The accompanying project page says LingBot-Touch was pretrained on 4,983 hours of recordings from piezoresistive, magnetic and visuotactile sensors. That data covers 464,431 interaction episodes and roughly 1.33 billion frames, giving the model a broader training base than a system tied to one tactile sensor or robot hand 1.
Robbyant also highlights a 5.46-million-parameter HaT-Tiny encoder for high-frequency downstream inference. In the company’s reported evaluation, the model was tested across 26 real-world dexterous-manipulation tasks on Sharpa Wave and XHand platforms, alongside 12 simulated tasks. Those figures come from the project and its technical report, rather than an independent audit 12.
The real-world examples include unlocking with a key, dispensing toothpaste and closing a resealable bag — tasks where pressure, contact and timing matter as much as a robot’s visual understanding. The project says frozen LingBot-Touch features improved the reported mean success rates of several policy families when added during post-training 1.
The release includes code and pretrained checkpoints through the project’s GitHub repository, while the technical report provides the underlying methods and evaluation details. The repository states that the project is available under the Apache 2.0 licence, subject to the terms of its own distribution 23.
Our opinion
LingBot-Touch is interesting because it treats touch as reusable experience rather than a bespoke add-on for one robot. The important test will be whether developers can reproduce the reported gains on hardware outside Robbyant’s evaluation setup. Open code and checkpoints make that question easier to investigate, but the headline numbers should still be read as project-reported results until independent teams test them.
For robotics researchers, the practical appeal is the attempt to bridge incompatible tactile inputs without throwing away the useful structure in each sensor. For everyone else, the release is a reminder that progress in robots is not only about bigger language or vision models: reliable manipulation also depends on knowing when something has been touched, squeezed, seated or moved.