Visuotactile Glossary A-Z — 30 Tactile-Sensing Terms Explained

A2 entries

Affordance: The set of actions an object permits a robot or agent to perform on it — e.g. a handle affords grasping, a button affords pressing. In visuotactile work, tactile data is often used to refine affordance estimates from vision once contact is made.
AnySkin: A 2024 magnetic-skin tactile sensor from NYU, designed to slip on like a fingertip cover and to remain calibration-stable across instances of the same robot hand. Open-source.

C2 entries

Compliance mech.: The mechanical property of yielding under load. A compliant fingertip absorbs small misalignments and impact forces; almost all visuotactile sensors are deliberately compliant by design (the gel deforms).
Contact-rich manipulation: The class of manipulation tasks where contact between the robot and the environment dominates the task dynamics — insertion, polishing, wiping, in-hand re-orientation. The natural application domain for visuotactile sensing.

D3 entries

DIGIT sensor: An open-source compact visuotactile sensor released by Meta AI Research in 2020, with full CAD, BOM and firmware under MIT license. Often the default reference sensor in 2020-2026 academic papers.
Diffusion policy: A policy-learning framework in which the action distribution of a robot is modelled by a denoising diffusion process. Introduced by Chi et al. (2023); extended in 2024 to consume visuotactile inputs alongside vision and proprioception.
Domain randomisation: A sim-to-real technique in which physical, visual or sensor parameters are randomised during training so that the resulting policy generalises to a wider range of real-world conditions. Particularly important for visuotactile sim-to-real.

E2 entries

Elastomer: A soft, elastic polymer — typically silicone or polyurethane — used as the deformable contact surface of a visuotactile sensor. The chosen elastomer chemistry determines hardness, lifetime and how cleanly the gel returns to its rest shape.
Embodiment: The specific physical body (and its sensors) of a particular robot. Cross-embodiment generalisation — a policy or model that works across different robot bodies — is one of the open problems in robotics learning, and a major motivation for tactile foundation models.

F2 entries

Force-torque sensor F/T: A six-axis sensor measuring the three-dimensional force and three-dimensional torque acting at a single point, usually mounted at the wrist of an industrial robot. The classical baseline for any contact-aware manipulation task.
Frame-rate: The number of tactile frames a sensor delivers per second. Photometric visuotactile sensors typically run 30 to 90 Hz; magnetic skins can reach kilohertz; force-torque sensors deliver scalar updates at ~1 kHz.

G1 entry

GelSight sensor / company: The original visuotactile design (Johnson and Adelson, MIT, 2009) and the spinout company that commercialises it. The lineage includes GelSight Mini (compact research sensor) and GelSight Wedge (edge-sensing variant).

H1 entry

Humanoid robot: A robot with a human-like body plan — two arms, two legs, a head — and increasingly with hands designed for human-tool use. Humanoid programmes are the largest near-term demand driver for visuotactile sensing.

I2 entries

Imitation learning: A policy-learning paradigm in which a robot policy is trained from human demonstrations rather than from explicit reward. Visuotactile feedback substantially improves imitation learning on contact-rich tasks.
In-hand manipulation: Re-orienting or re-positioning an object between the fingers of a single hand without re-grasping. A canonical hard problem for vision-only systems and a flagship visuotactile demonstration since 2021.

M3 entries

Magnetic skin: A tactile sensor in which a magnetised elastomer is read by an underlying magnetometer array, with no internal camera. Trades spatial resolution for high frame-rate and easy field replacement. Examples: ReSkin, AnySkin.
Marker tracking: A visuotactile sensing approach in which the internal camera tracks a discrete grid of physical markers embedded in or on the gel, rather than reconstructing a continuous depth field. Used by TacTip and several GelSight variants.
MPPI control: Model Predictive Path Integral control. A sampling-based predictive controller often used as a baseline in tactile-feedback control, especially for contact-rich tasks where gradient-based MPC is poorly conditioned.

P3 entries

Peg-in-hole: The textbook contact-rich assembly task: aligning and inserting a peg into a hole with sub-millimetre clearance. The benchmark on which most visuotactile insertion work has been demonstrated.
Photometric stereo: An optical technique in which a surface is illuminated from several known directions and its per-pixel surface normals (and from there its depth) are reconstructed from the resulting shading. The core principle behind GelSight-family sensors.
Pinch point: The geometric region of contact between a robot fingertip and an object; informally, the tactile equivalent of a camera's footprint. The size and shape of the pinch point is what limits how much information a single tactile reading can carry.

R1 entry

ReSkin: A 2021 magnetic-skin tactile sensor from Meta AI Research and CMU, designed for low cost and easy replacement of the worn elastomer layer.

S3 entries

Sim-to-real: The problem of transferring a policy trained in simulation to a real robot. Particularly hard for visuotactile data because simulating the full optical-mechanical behaviour of a soft gel is computationally expensive.
Slip detection: The real-time detection of micro-motion of an object relative to the gripping fingers. The earliest practical demonstration of why visuotactile sensors do something a force-torque sensor cannot.
Soft-Bubble: A whole-finger compliant tactile sensor design from Toyota Research Institute, built around an internal camera observing an inflated transparent membrane. Used for fragile-grasp and large-area contact tasks.

T3 entries

TacTip: A family of marker-tracking visuotactile sensors developed at the Bristol Robotics Lab, biomimetically modelled on the structure of human skin papillae.
Tactile foundation model: A neural-network model trained on tactile data from many sensors and tasks, intended to provide a general-purpose tactile representation. Several proposals appeared in 2025 and 2026; full maturity is still an open question.
Tactile-RL: Reinforcement-learning approaches that include tactile inputs in the policy observation space. Distinct from imitation learning in that the policy is improved through trial and error, with reward typically driven by task success.

V1 entry

Visuotactile sensor: A tactile sensor that uses an internal camera to observe the deformation of a soft contact surface, producing an image-like tactile signal at video frame rates.

W1 entry

Whole-hand manipulation: Manipulation tasks that exploit the full hand surface rather than only the fingertips — cradling a ball in the palm, rolling an object across the back of the fingers. Soft-Bubble-style large-area tactile sensors are well suited to this regime.

Y1 entry

YCB-Tactile: A tactile-data extension of the well-known YCB Object and Model Set (Yale-CMU-Berkeley), providing reference visuotactile recordings of the standardised YCB objects across multiple sensors. Used for cross-sensor benchmarking.

Sensors League Table

Side-by-side comparison of all the sensors named in this glossary — GelSight, DIGIT, ReSkin, AnySkin, Soft-Bubble, TacTip.

Open dossier →

Applications

Where in-hand manipulation, slip detection, peg-in-hole and the rest of the use cases actually live.

Open dossier →

Research Timeline

The papers behind the terminology — GelSight (2009) through tactile foundation models (2026).

Open dossier →

Glossary — the vocabulary of visuotactile sensing.

Sensors League Table

Applications

Research Timeline