Robbyant releases open-source AI to help robotic hands reuse their sense of touch

Changing a robotic hand’s sensors need not mean rebuilding all its touch perception. That is the promise behind Robbyant’s release of LingBot-Touch, an open-source AI model that translates different tactile signals into a shared format. The company reports improved manipulation results across two robotic hands and four models that control their actions.

Robbyant, an embodied AI company within Ant Group, has made the model available through Hugging Face and GitHub, alongside code and a technical report. Its researchers posted their paper on November 9, 2026. The release provides a perception layer between touch sensors and a robot’s controller, rather than a replacement for the controller itself.

Different sensors, one description of contact

The problem starts with incompatible inputs. The same physical contact can produce different signals depending on how a sensor works and where its contact points sit. A model built around one arrangement may therefore require fresh data and retraining when developers change the sensor or hand. Touch hardware alone does not solve that software problem.

LingBot-Touch uses separate processing paths for different sensing technologies, then maps their contact points into a common fingertip coordinate system. A shared model learns where contact occurs, how force is distributed and how both change over time. It packages that information into tactile tokens: numerical representations that an action model can use to choose movements.

Robbyant says pre-training used nearly 5,000 hours of real-world tactile data, including magnetic, visual-tactile and piezoresistive sensing. It combined data from human-worn tactile gloves and remotely operated robotic hands. The researchers say the system can also support new sensor types.

A bottle cap separates sensing from interpretation

Robbyant evaluated the model in 26 real-world tests. Across four action models, it reported average task-success improvements of approximately 14 to 26 percentage points compared with operation without tactile input. Those are company-reported results for the tested configurations, not a guarantee of the same improvement on another robot.

• Hardware: the Sharpa Wave and XHand dexterous robotic hands.
• Controllers: LingBot-VLA, π0.5, Diffusion Policy and T-Rex, spanning several approaches to choosing robot actions.

One bottle-cap experiment offers a more revealing comparison than touch versus no touch alone. Using T-Rex to control Sharpa Wave, the team compared operation without tactile feedback, direct use of raw sensor signals and use of LingBot-Touch’s processed representation. The reported gap suggests that interpreting the signals mattered, not simply supplying more input.

56.25%Without tactile input

Robbyant reported this success rate for T-Rex controlling Sharpa Wave without tactile feedback.

62.5%Raw sensor signals

Directly feeding raw tactile signals into the controller produced this company-reported result.

90%LingBot-Touch representation

Using LingBot-Touch’s processed tactile representation produced this company-reported result.

Reusing perception still leaves a controller to train

The XHand experiment addressed a different question: whether learned touch perception could transfer to hardware absent from pre-training. According to Robbyant, XHand’s magnetic sensor was not in that dataset. All the model’s prior magnetic-touch experience came from data collected with gloves worn by humans.

The team kept LingBot-Touch’s learned parameters unchanged while training only the action model on XHand demonstrations. That distinction limits what “reuse” means here. The experiment reused the touch-perception component; it did not eliminate training for the controller that turns perception into movement.

All 26 physical tests used LingBot-Touch-Tiny, with about 5.3 million parameters. Robbyant says pairing it with LingBot-VLA, π0.5 and T-Rex increased total model size by about 0.13% to 0.16%. A larger version showed further potential in perception tests, but the physical-task results described here are for Tiny.

Developers can use LingBot-Touch with controllers outside Robbyant’s own model family. The practical offer is a reusable touch-processing layer while robotics teams retain their choice of action model.