There is no single “best” visuotactile sensor. The right choice depends on whether the task is dexterous in-hand manipulation, fragile-grasp slip detection, surgical tool feedback or production-grade humanoid hand integration. The table below is intentionally flat — no rankings, no scoring — so that the trade-offs between modality, resolution, frame-rate and openness can be read directly. All values are taken from the publicly available datasheets, papers or repositories of each sensor and labelled as such where they are estimates.
Comparison TableSensors at a Glance — 2026 Edition
| # | Sensor | Origin | Modality | Resolution | Frame-rate | Form factor | Licensing | Year |
|---|---|---|---|---|---|---|---|---|
| 1 | GelSight Mini Most-used commercial visuotactile |
GelSight Inc. (USA) | Photometric stereo, RGB camera + elastomer | ~25 µm / 320x240 | ~30 fps | Compact18x24 mm sensing area | Commercial | 2022 |
| 2 | GelSight Wedge Edge / contour sensing variant |
GelSight Inc. (USA) | Photometric stereo, wedge geometry | ~50 µm / 640x480 | ~30 fps | Wedgeangled sensing surface | Commercial | 2021 |
| 3 | DIGIT Open-source compact reference |
Meta AI Research (USA) | Photometric stereo, mini camera | ~50 µm / 320x240 | ~60 fps | Fingertip~20x27 mm sensing | Open Source | 2020 |
| 4 | DIGIT 360 Multi-modal successor |
Meta AI Research (USA) | Photometric + audio + thermal (reported) | ~7 µm / multi-modal | up to ~120 fps | Sphericalfingertip dome | Open Source | 2024 |
| 5 | ReSkin Magnetic-skin tactile sensor |
Meta + CMU (USA) | Magnetometer array under magnetised elastomer | ~1 mm / sparse field | up to 1 kHz | Skin patchswappable, low cost | Open Source | 2021 |
| 6 | AnySkin Slip-on, calibration-free magnetic skin |
NYU (USA) | Magnetic, designed for cross-instance generalisation | ~1 mm / sparse field | several hundred Hz | Slip-onfield-replaceable skin | Open Source | 2024 |
| 7 | Soft-Bubble Whole-finger compliant sensor |
Toyota Research Institute (USA) | Internal camera, inflated transparent membrane | ~1 mm / depth field | ~30 fps | Inflated bubblelarge compliant surface | Research | 2020 |
| 8 | TacTip Marker-tracking biomimetic sensor |
Bristol Robotics Lab (UK) | Internal camera tracking embedded marker pins | ~100 µm / marker grid | ~50-90 fps | Domebiomimetic papillae structure | Open + Commercial | 2009 / 2018 redesign |
| 9 | 6-DoF F/T sensor e.g. ATI Nano17, Bota SensONE |
ATI / Bota / Robotous | Strain-gauge force-torque, mounted at wrist or fingertip | 6 scalar values | ~1 kHz | Cylindrical20-50 mm diameter | Industrial | established |
“The single most important variable when choosing a tactile sensor is not resolution — it is whether your control loop can actually consume an image stream at 60 frames per second.”
— Common observation in 2025-2026 manipulation papersHow the Field Divides — Three Modality Families
Camera + Elastomer
The original GelSight lineage. An internal camera observes a deformable gel illuminated from the side; per-pixel depth is reconstructed photometrically. Highest geometric resolution, moderate frame-rate, larger fingertip volume. Members: GelSight Mini, GelSight Wedge, DIGIT, DIGIT 360.
Camera + Marker Pins
An internal camera tracks an array of physical markers embedded in or on the gel. Direct measurement of shear forces and slip, lower geometric precision than photometric. Members: TacTip, GelSight variants with marker overlays, several Chinese university designs.
Magnetometer + Magnetised Skin
No camera at all: an array of magnetometers reads the displacement of a magnetised elastomer skin. Higher frame-rate (kilohertz possible), much lower spatial resolution, easy field-replacement. Members: ReSkin, AnySkin, several university one-offs.
Notes on Methodology
Resolution and frame-rate values are taken from the most recent publicly available datasheet or paper for each sensor. Where the design has multiple operating modes (DIGIT 360 in particular), the highest commonly reported value is used. The licensing column simplifies a more nuanced reality: most “open source” sensors release CAD, BOM and firmware under permissive licenses, but assembled units are sold by partner vendors. Prices are not included because publicly disclosed prices vary by quantity, region and integration support, and are generally negotiated.
The GelSight family (entries 1-2) has the highest spatial resolution and the largest installed base in industry. The DIGIT family (entries 3-4) is the dominant open-source reference in academic research. ReSkin and AnySkin (entries 5-6) trade tactile-image richness for raw frame-rate and replaceability. Soft-Bubble (entry 7) targets a different niche: large compliant contact area for whole-hand tasks. TacTip (entry 8) remains the most common marker-based design, particularly in European labs. The 6-DoF force-torque sensor (entry 9) is included to remind that for many manipulation tasks — particularly assembly — a wrist-mounted F/T is still the right tool.
FAQFrequently Asked Questions
Which visuotactile sensor has the highest spatial resolution?
Among publicly documented designs, GelSight Mini reports an effective tactile resolution down to roughly 25 micrometres on a sensing surface of about 18 by 24 millimetres. Photometric-stereo sensors generally outperform marker-tracking and magnetic skins on raw geometric resolution because they reconstruct a per-pixel depth map rather than tracking a finite number of features.
Is DIGIT really fully open-source?
DIGIT was released under an MIT license by Meta AI Research in 2020, including CAD files, BOM and reference firmware. The DIGIT 360 successor, announced in 2024, also publishes mechanical design and firmware. Hardware can be self-built or purchased from partner vendors. The dataset stack and high-level libraries (PyTouch, Tacto simulator) are likewise open source under permissive licenses.
Why is a 6-DoF force-torque sensor in a visuotactile comparison?
Six-axis force-torque sensors mounted at the wrist are the long-standing baseline for any contact-rich manipulation task. They are not visuotactile, but every benchmark or product comparison should include them as the reference point that visuotactile sensors must improve on. Industrial F/T sensors from ATI Industrial Automation, Robotous and Bota Systems are the de facto standard.
Are these sensors available off-the-shelf?
GelSight Mini and GelSight Wedge are commercially available from GelSight Inc. TacTip is sold via the Bristol Robotics Lab and partner integrators. DIGIT, ReSkin and AnySkin are open-source designs that require self-assembly or sourcing through partner vendors. Soft-Bubble was developed at Toyota Research Institute and is not currently sold as a standalone product.
What does sensor frame-rate actually limit?
Frame-rate determines how fast a control loop can react to a tactile event. Slip detection during a fragile grasp typically needs at least 30 to 60 Hz to be useful; assembly insertion benefits from 90 Hz or higher. Magnetic skins like ReSkin and AnySkin can run at kilohertz rates but with much lower spatial resolution. The right trade-off is task-specific.
How quickly do visuotactile gels wear out?
Publicly reported lifetimes vary widely. GelSight Mini gels are typically rated for several thousand contacts under nominal load before noticeable surface wear; ReSkin advertises easy field replacement of its magnetic skin layer. Real-world durability depends heavily on contact force, abrasiveness of objects and exposure to oils. Most production-grade humanoid programmes plan for routine fingertip replacement.