Tactile Data

Pressure, shear force, and contact-surface deformation recorded by a tactile sensor when it touches an object.

Tactile data comes from tactile sensors mounted on a fingertip, palm, or gripper, recording contact location, normal force (how hard something is pressed), shear force (related to slipping), and how the contact surface deforms. Roughly two kinds of sensors exist: visuotactile sensors (such as GelSight and DIGIT) use a built-in camera to photograph the deformation of an elastic gel pad and output something like a photograph, a tactile image; array-style sensors — piezoresistive, capacitive, or magnetic — output a numeric reading per taxel. When the hand blocks the camera's view, or force needs to be controlled precisely (peg insertion, unscrewing a cap, holding something fragile), touch supplies information a camera can't. The difficulty is that sensor models vary and their data doesn't transfer between them, and ground-truth force and slip are hard to label, so Meta's Sparsh used self-supervised pretraining on more than 460,000 unlabeled tactile images. Datasets such as RoboMIND 2.0 now record touch in sync with vision and joint state.

ExamplePinching a strawberry with a GelSight-equipped gripper, the size of the contact area and the displacement of the tactile image's surface markers reveal how tightly it's being held and whether it has started to slip.

Also called

Visuo-tactile Data

Related

Tactile Sensor · Vision-Based Tactile Sensor · Tactile Representation Learning · Visuo-Tactile Fusion · Multimodal Data · Slip Detection

Sources

Sparsh: Self-supervised Touch Representations for Vision-based Tactile Sensing
Touch and Go: Learning from Human-Collected Vision and Touch
RoboMIND 2.0 (arXiv HTML)

As of

2025-12