Robbyant, the embodied-AI company under Ant Group, released LingBot-Touch, a tactile foundation model for dexterous robotic manipulation. The project can translate signals from different tactile sensor types into a common representation that a robot’s action model can use. Robbyant says the model was pretrained on nearly 5,000 hours of recordings and supports magnetic, piezoresistive and visuotactile sensing. The release extends the company’s LingBot family beyond visual and spatial perception by focusing on contact information: where a robot touches an object, how much force it applies and how that contact changes over time. TechNode’s report covers the release timing and the project’s publication status.
A shared representation aims to bridge incompatible tactile sensors
Different tactile sensors produce different kinds of data. Magnetic sensors can measure changes associated with force, piezoresistive sensors respond to pressure, and visuotactile sensors use visual deformation to infer contact. Models trained on one sensor may not transfer cleanly to another because the signals and physical layouts differ. LingBot-Touch uses modality-specific tokenizers and a shared Haptic changeer to map these signals into a common latent representation. That design aims to let an action model reuse tactile experience across different hands and sensors rather than requiring a separate training pipeline for every hardware configuration.
Robbyant’s technical material describes a corpus of 4,983 hours, more than 464,000 interaction episodes and approximately 1.33 billion tactile frames. The smallest encoder is listed at about 5.46 million parameters. The company reports evaluation across 26 real-world tasks on two dexterous robotic hands, with the tactile representation improving the average success rate of four action models. The reported gains are company results and have not yet been independently replicated across unrelated laboratories or hardware. The task set and selected robot hands matter when interpreting the figures, particularly for claims about transfer to unseen sensors.
The published weights do not make every part of the system open
The project page provides model information and links to code and model repositories. The release is useful to researchers because it makes an inference component available for experimentation, but it should not be described as a fully open training stack without checking each repository. Model weights, inference code, training data, task-specific heads and evaluation procedures can be released separately. A team intending to reproduce the work needs to verify which components are available, what licence applies to each and whether the hardware adapters support its own sensors. A model can be open for inference while important parts of the training pipeline remain unavailable.
Tactile perception is important for tasks where vision alone cannot reliably estimate contact force or object deformation. A robot may see a bottle cap but still need touch feedback to determine whether it has gripped too tightly or whether the cap has begun to turn. Similar problems arise when handling soft packaging, inserting connectors or manipulating fragile objects. LingBot-Touch provides a way to standardise tactile inputs for downstream policies, but practical adoption will depend on latency, calibration, sensor compatibility and robustness to wear. Independent tests on different hands and tasks will determine whether the shared representation becomes a useful robotics building block.
The model’s practical value will depend on how reliably it transfers across sensors that were not part of its training data. A robot hand can have different fingertip geometry, sensor placement and calibration, all of which affect the signals received by the model. A shared representation can reduce the amount of adaptation required, but developers still need to validate it against their hardware and tasks. The technical report’s results should be reproduced by independent groups using different objects, hands and operating conditions. It is also important to understand the release boundaries: a frozen encoder can be useful for inference, while retraining or adapting the full model may require access to components that are not included in the public release.