Sensor evidence / Commercial optical touch
GelSight Mini: specifications and a usable research workflow
Review GelSight Mini specifications, RGB and 3D workflows, calibration limits, replaceable gel, Python examples, and evidence for robotics use.
Published 2026-09-12 | Updated 2026-09-16 by RoboSkin.ai Editorial Team
- 6
- sections
- 4
- questions
- 6
- next routes
Short answer
What you need to know
- 1
GelSight Mini is a commercial optical tactile sensor. A camera observes an illuminated compliant gel as it conforms to a contacted surface. The official datasheet reports an 18.6 × 14.3 mm field of view, an 8 MP camera, and a 25 FPS camera frame rate.
- 2
Those specifications are distinct from reconstruction settings: the reviewed gsrobotics examples use a configurable image-processing pipeline, with a 320 × 240 default configuration shown in the README. A sensor specification, a processed image size, and the spatial accuracy of a learned model are different measurements.
- 3
The sources below provide a path from live RGB capture to example 3D reconstruction. RoboSkin.ai reviewed the manufacturer documents and code documentation; we have not acquired a sensor or independently reproduced its performance.
Topic 01
Manufacturer specifications with operating conditions
The following values are transcribed from page 3 of the official datasheet reviewed on September 12, 2026. They belong to GelSight Mini, and should not be copied onto other GelSight products or all optical tactile sensors.
| Property | Manufacturer-reported value | Interpretation |
|---|---|---|
| Field of view | 18.6 mm horizontal × 14.3 mm vertical | Sensing view, not overall enclosure dimensions |
| Camera | 8 MP; 25 FPS | Separate camera specifications; check the actual negotiated stream mode |
| Gel thickness | 4.25 ± 0.20 mm | Gel thickness is not sensor thickness |
| Illumination / gel | RGB LEDs; Lambertian silicone gel | Optical reconstruction depends on the gel and illumination state |
| Preferred operating temperature | 0–30 °C | Datasheet environmental conditions |
| Storage / humidity | −25 to +60 °C; up to 80% RH, non-condensing | Storage range does not establish the same range for operation |
| Gel durability | 1,000 coin presses | A named manufacturer test; not a general industrial cycle rating |
Topic 02
Start with image capture, then inspect the 3D pipeline
GelSight links a browser application for an initial live view. For programmable work, its gsrobotics repository documents live-view, dual-camera, 3D-viewer, and basic OpenCV demos, with installation instructions for Windows and Linux. The README also links case and gripper-adapter CAD models.
A useful first session identifies the camera, captures an unloaded gel image, records a short contact sequence, and checks the saved frames before adding a robot controller. The 3D demo uses a supplied model and configuration for reconstruction. Preserve the model file, preprocessing, image dimensions, scale, and marker-handling settings with each experiment.
The datasheet mentions ROS and ROS 2, including older distributions. Treat that as documented historical compatibility. Verify a maintained driver and dependency set for the distribution and operating system you actually use; this page does not certify every current ROS release.
Topic 03
Calibration, force, and replacement gel
The official repository describes example reconstruction from RGB images and includes a reference calibrated model. The existence of that example does not establish a calibrated force vector for every purchased unit. For metric geometry, verify the image scale and reference depth on your device; for force, collect synchronized reference-force measurements.
The cartridge is described as user-replaceable without tools. After a change, inspect unloaded appearance, focus, contact coverage, and reconstruction on reference shapes. Replacement can alter the image distribution seen by a model, even when the mechanical cartridge returns to a similar position.
Marker motion can support shear estimation, but a visible marker displacement is not automatically a value in newtons. Record the calibration procedure and quantify its error independently from grasp or teleoperation success.
Topic 04
Reading robotics evidence correctly
RoboSkin.ai covers studies using GelSight Mini in different roles. The spatial-tactile teleoperation article examines a specific mapping, display, apparatus, and human task. The SoftVTBench article distinguishes simulated visuo-tactile observations from a physical collection claim. Their outcomes should remain attached to those setups.
When comparing a new policy with a published result, match the robot, sensor revision, gel, input processing, test objects, split, and success criterion. Use the experiment evidence module as a model for retaining these fields with each numerical result.
Topic 05
Access and reproducibility checklist
Hardware is commercially offered by GelSight; use the manufacturer for current price, stock, and support. The reviewed gsrobotics repository is public and carries a GPL-3.0 license. That repository license does not automatically cover separately linked datasets or every third-party model.
Before committing to a robot integration, check the mounting envelope and cable path, establish a repeatable capture configuration, validate the metric you need, and test the effect of gel replacement. Keep the official datasheet and a pinned software revision with the resulting experiment record.
Topic 06
Prepare a calibration record before interpreting contact
The calibration workflow separates no-contact image baselines, reference geometry for depth reconstruction and independently measured force targets. Use the blank record to retain units, repeats, environmental conditions and invalid samples. The workflow is source-based; no new hardware measurements are claimed.
Common questions
FAQ for this topic
What is the GelSight Mini sensing area?
The reviewed manufacturer datasheet gives a field of view of 18.6 mm horizontally by 14.3 mm vertically. This is different from the enclosure dimensions and from an image-processing crop.
Does GelSight Mini provide RGB, depth, or force?
The camera provides RGB observations. Official examples process these into 3D representations using a model. Calibrated force and task-specific accuracy require additional evidence; they should not be assumed from the live image or the datasheet.
Why do the specifications say 8 MP while examples use 320 × 240?
The camera specification and a demo processing configuration describe different stages of the system. Record the actual capture format and reconstruction resolution, rather than treating either number as a force or depth accuracy measurement.
Can the gel be replaced?
The manufacturer describes a tool-free, user-replaceable cartridge. Check the image reference and task calibration after replacement; mechanical replacement alone does not verify that a learned model transfers unchanged.