Tactile AI: touch data for Physical AI

Three recent studies make different engineering decisions: detect an event for a separate controller, learn joint actions directly from tactile history, or render contact into a visual policy. Choose the decision your robot needs to make before comparing reported accuracy or success. The linked reviews use the original v1 papers and retain the training, timing and hardware limits.

Choose one output before choosing a model: a contact classification, an estimated physical quantity, a predicted tactile observation, or a robot action. Specify how an independent reference or a repeated task will tell you whether that output is useful. A good first experiment can be small, provided its measurement and success criterion are explicit.

For a hardware experiment, the DIGIT, GelSight Mini, and ReSkin guides document three different acquisition paths. For an experiment with existing data, the tactile dataset directory exposes the sensor, task, access, and split information to check before training. Use the experimental evidence module to see how a result changes meaning when its sample, comparator, or protocol changes.

A tactile AI stack starts with a contact surface and ends with an action or measurement loop. Between those endpoints, the system needs sensing materials, electronics, timestamps, calibration, feature extraction, model inputs, and robot middleware.

If the robot cannot use the signal in a control or evaluation loop, the system is only collecting touch data. Tactile AI begins when that data changes what the robot can decide or verify.

Humanoid robots and dexterous hands operate in contact-rich settings. Vision can guide the robot toward an object, but a hand often blocks the camera once grasping begins.

Touch data can reveal whether an object is seated correctly, sliding, deforming, or being squeezed too hard. That information matters for household tasks, warehouse handling, prosthetics, assistive devices, and research platforms.

Touch is not limited to the hand. Plantar pressure arrays can expose the support realized beneath a humanoid foot after touchdown, giving a locomotion policy spatial evidence about partial, asymmetric, compliant, or shifting contact.

Tactile AI claims should be tied to measured tasks. A demo that classifies contact on a benchtop is different from a robot hand that adjusts grip during motion.

Useful validation includes sensor drift, response time, synchronization with joint state, robustness after mounting, and whether the tactile signal improves a real robot behavior.

Tactile sensing is the measurement layer. Tactile AI is the larger perception-and-action system that turns those measurements into a representation, prediction, decision, or controller input. Keeping the boundary clear prevents a sensitive sensor demo from being described as an intelligent robot system without task evidence.

The operational chain is contact → sensing → calibrated and timestamped data → tactile representation → model inference → controller or policy → robot action → measured outcome. Each transition has a contract: units, coordinate frame, sampling rate, latency, uncertainty, and failure behavior.

A robust system keeps measured values separate from inferred values. For example, a pressure array may measure taxel response while a model estimates slip risk; the controller then decides whether to increase grip, regrasp, slow the motion, or stop. Logging all three levels makes the result auditable.

A vision-language-action model can provide semantic task context and propose actions, while tactile feedback supplies local physical evidence after contact. A tactile or visuo-tactile world model instead predicts how contact state may evolve under a candidate action. These roles can be combined, but a VLA label does not prove high-frequency touch control and a plausible world-model rollout does not prove safe execution.

Physical AI is the broader embodied system: vision observes the scene, language represents goals and knowledge, proprioception describes the robot body, and touch grounds the interaction at the contact surface. Tactile AI is the part of that system responsible for interpreting and using touch.

The systems below solve different parts of the stack and should not be collapsed into one leaderboard. The source status and hardware contract matter as much as a reported metric.

RoboSkin.ai tracks entities through their public research assets instead of presenting a vendor ranking. This keeps company, laboratory, sensor, and robot relationships traceable to primary sources.

Robotic tactile sensing supplies observations of contact; tactile AI depends on a defined representation, calibration target and data-quality contract. Keep raw image or sensor units separate from derived depth and force, and retain unknown values through preprocessing.

Start with a sensor-specific calibration plan, practice missing-data checks in the synthetic Python exercise, then inspect episode and timestamp conventions before combining observations with robot actions. None of these data exercises establishes physical sensor accuracy.

Two September 2026 preprints expose different data choices for tactile AI. DexTouch-WM transfers human glove observations into an action-conditioned robot world model using a shared tactile layout and motion retargeting. Bench2Dex provides a simulation pipeline for bimanual demonstrations, geometric tactile maps, and controlled policy evaluation.

Keep their evidence separate: better predicted contact does not guarantee better policies trained on synthetic observations, and a simulated tactile image is not a physical force measurement. Our September 18 source reviews examine the result definitions, data contracts, and release status before connecting either project to a training workflow.