Tactile engineering guide
Tactile Sensor Calibration: DIGIT and GelSight Mini Workflows
Plan DIGIT and GelSight Mini calibration, held-out validation, error and uncertainty checks. Know when to reuse or refit, with blank recording worksheets.
Tactile sensor calibration starts with a target quantity. A stable reference image supports image comparison; geometry calibration supports depth reconstruction; force estimation needs force reference measurements and its own validation.
DIGIT and GelSight Mini are vision-based tactile sensors. Their camera images and depth demonstrations should not be relabeled as calibrated newtons or pressure.
Official sensor materials, source READMEs and the 3D Cal paper were checked on September 16, 2026. This is a source-based workflow and a blank recording template. RoboSkin has not run DIGIT, GelSight Mini, a calibration rig, TouchNet training or a force experiment for this guide. No measured accuracy or hardware compatibility is claimed. The November 3, 2026 extension reviews calibration validation and NIST uncertainty guidance; its scenarios and arithmetic are illustrative and no physical experiment was added.
Choose one calibration target at a time
In robotic tactile sensing, calibration links a sensor output to a defined reference under stated conditions. Decide the target, unit, coordinate frame and operating range before collecting examples. A model can be useful for geometry without being a force sensor.
| Target | Reference and procedure | What it does not establish |
|---|---|---|
| Image baseline | Record repeated no-contact frames at fixed illumination, exposure, focus, crop and resolution. Characterize drift and noise. | RGB differences alone do not identify depth in millimetres or force in newtons. |
| Geometry / depth | Use known probe geometry and controlled displacement/position labels; fit or assess an image-to-gradient/depth mapping. | Depth reconstruction is not a calibrated normal/shear force model. |
| Force estimation | Collect synchronized, independently measured normal/shear loads with a suitable load cell or force/torque reference across relevant contact conditions. | A depth model, gel stiffness assumption or uncalibrated motor current does not supply this reference. |
What the DIGIT and GelSight Mini sources support
The original DIGIT interface documents connecting by serial number, reading camera frames and selecting supported stream settings. The reviewed README specifies a default VGA 640 × 480 stream at 30 fps. Those defaults describe the interface, not a per-device calibration certificate. This guide concerns original DIGIT, not DIGIT 360 or other later sensors.
GelSight Mini’s official gsrobotics examples cover live image viewing, recording, marker tracking, depth estimation and point clouds. Pixel-to-millimetre conversion depends on the documented model and image resolution; do not transfer a constant to a different crop, lens, gel or reconstruction. Marker displacement is a useful observation but is not force in newtons without a corresponding reference calibration.
The 3D Cal paper (arXiv:2511.03078v1) demonstrates automated probing of DIGIT and GelSight Mini with a repurposed FDM printer. TouchNet predicts surface gradients, which are integrated into depth maps. The paper explicitly describes force-sensor integration for normal and shear targets as future work. Its reconstruction results are the authors’ experiments; no numbers here are a RoboSkin reproduction.
Prepare the sensor, reference and record
For either sensor, record model, serial number, gel identity/condition, mount, camera settings, software revision and an unmodified raw image stream. Use a repeatable mount and confirm the field of view. Save a new baseline after changes to gel, optics, illumination, exposure or geometry. Retain the previous calibration instead of silently overwriting it.
For depth work, prepare an independently characterized probe and positioning reference, define the sensor coordinate system and retain uncertainties in probe size, alignment and displacement. A commanded printer position is not automatically the true indentation depth. The 3D Cal authors used a spherical probe and a purpose-built mount; reproduce only a workflow appropriate to your actual equipment and device limits.
For force work, add a reference transducer with a known range, unit and calibration record. Define axis signs, normal/shear components, loading and unloading directions, dwell times and synchronization. No force procedure can be completed from the supplied blank template alone.
If the sensing surface needs attention first, use the gel inspection and replacement guide before recording the new calibration baseline.
DIGIT workflow: baseline first, then a geometry model
1. Identify the original DIGIT by serial number and inspect the raw stream using its official interface. Record the chosen resolution/FPS and available illumination/exposure controls. Fix configurable image settings for the run; if a control is unavailable or automatic, record that limitation.
2. With no contact, capture a short sequence after the image stabilizes. Save the baseline images and summarize per-channel mean, temporal variation and saturated pixels in a defined region. Choose an acceptance threshold from the intended task; this guide supplies no universal noise threshold.
3. For geometry calibration, collect contacts at a planned spread of surface locations and reference indentations with a known probe. Include independent repeated load/unload cycles and baseline checks between groups. Record reference uncertainty, probe placement and invalid captures. Do not count adjacent video frames as independent experimental repeats.
4. Fit or fine-tune an appropriate reconstruction using the selected source workflow, holding out locations and contact sessions. Compare reconstructed depth against independently defined geometry on held-out contacts and objects. Keep the original images, reference labels and model revision together; no training command or device movement is executed by this guide.
GelSight Mini workflow: record the gel and reconstruction settings
1. Identify the Mini, gel type and whether the surface carries markers. Confirm image capture with the official live-view example before assessing depth. Record the raw resolution, processing resolution, crop and any marker-removal settings.
2. Capture stable no-contact references in the same configuration used for contacts. Inspect uneven illumination, saturation, contamination and gel damage. Baseline adjustment can improve image consistency; it does not validate a metric reconstruction.
3. Use the official depth/point-cloud example to inspect the reconstruction interface, then evaluate with a known geometry and independently defined spatial scale. The demo is a starting point for verification. If using 3D Cal, keep its sensor-specific acquisition, training and hold-out records rather than assuming a downloaded model transfers unchanged to your Mini.
4. For marker gels, retain marker positions and tracking failures as separate signals. If a task needs force, acquire synchronized reference load data and evaluate a separate force model. A marker motion plot or a plausible point cloud is not evidence of force accuracy.
Download a calibration record template
Use one row per reference comparison at a contact location, repeat and loading phase. A header-only CSV is provided so no placeholder values can be mistaken for measurements. The accompanying field dictionary groups identification, units, references, repeated trials, environment, clocks and evidence paths. Both files are CC0-1.0.
Record a reference value and uncertainty with its unit; an estimate and its unit; planned repeat count and actual repeat index; temperature, humidity, gel condition and camera settings. Preserve sample and receive timestamps with their clock domains and a synchronization method. Leave unavailable measurements blank with an explanation; do not fill missing references with zero.
Set evidence_type to hardware_measurement, synthetic_demo or source_note. Keep these in separate runs and reports. This download contains no synthetic or hardware measurements; the Python and LeRobot exercises elsewhere use explicitly synthetic data and cannot establish a device calibration.
Design the validation split around the claim
A calibration fit answers whether a model explains its calibration examples. Validation asks whether it works on the contacts you intend to use. Choose the deployment claim first, then reserve entire physical contacts and sessions for evaluation. Splitting neighboring frames from one indentation between training and test sets can make the test much easier than a new contact.
The matrix below is a proposed test design. Change one factor at a time to diagnose a failure, then test relevant combinations to assess the installed system. Keep the held-out partition unchanged while selecting the model. If a failed evaluation is used to tune the model, create a fresh evaluation set before claiming transfer.
| Proposed claim | Hold out or vary | What to report separately |
|---|---|---|
| New contacts on the same mounted sensor | Whole contacts and acquisition sessions, including repeated unloaded baselines. | Contact count, session count, invalid samples and residuals by location. |
| Performance across the sensing surface | Unseen positions including edges; independently specified reference geometry. | Spatial error map and contact-region errors, not only an image-wide average. |
| Different object geometry or loading direction | Probe/object geometries, orientations and normal/shear load combinations. | Within-range versus out-of-range results; normal and shear errors in their defined axes. |
| Stable behavior during a held contact | Dwell duration, loading/unloading phase and relevant temperature conditions. | Change at constant reference load, baseline drift and history dependence. |
| Transfer after remounting or pad replacement | A separate installation or surface identity with its own reference checks. | Before/after results and whether the old calibration, a new offset or a refit was used. |
Evaluate error, repeatability and drift
Use held-out contacts and sessions. For each valid matched reference, calculate residual = estimate − reference in the same unit and coordinate convention. Report bias (mean residual), MAE (mean absolute residual) and RMSE (square root of mean squared residual), with sample counts, invalid counts and reference uncertainty. Do not pool millimetres, newtons and pixels into one error metric.
For depth maps, report the evaluated region, alignment method and masking rules; separate contact and non-contact regions so a large flat background cannot hide contact errors. Examine spatial error across the sensing surface and performance on held-out geometries. For forces, report normal and shear axes separately, plus load-range coverage, hysteresis and held-out sessions.
Repeatability requires independent contacts at matched conditions. Compare repeated load/unload cycles and baseline measurements across time, temperature and gel condition. Report drift and failed measurements. Pick acceptance criteria for the application before testing; no sensor accuracy, threshold or minimum sample count is established by this guide.
Read bias, spread and failure rate together
Illustrative arithmetic, not sensor measurements: residuals of −0.4, −0.4, +0.4 and +0.4 N have zero mean bias but MAE and RMSE of 0.4 N. A mean near zero can hide substantial errors of opposite sign. Also inspect errors by load, position and loading phase; a single overall metric can hide the condition that matters most.
Report how many attempted contacts produced valid estimates and why the rest were excluded. Percentiles need their sample count and aggregation rule: a per-frame percentile weights long recordings more heavily than a per-contact summary. Neither a small observed maximum nor a high percentile proves a worst-case bound for future contacts.
Do not confuse residual error with measurement uncertainty
A residual compares two estimates: the sensor output and its reference. A small residual is not proof that either is close to the unknown true value. Record uncertainty in the reference load or geometry, positioning and alignment, synchronization and relevant environmental effects. Repeatability alone does not include every contribution.
NIST TN 1297 classifies uncertainty components by how they are evaluated: statistical analysis or other evidence. These categories are not synonyms for random and systematic error. For an uncertainty budget, define the measurement model and express each contribution in compatible output units through its sensitivity. Correlations matter; do not simply add a reference certificate value, an MAE and a repeatability number.
Record the coverage factor and assumptions when using expanded uncertainty. A reported interval is not automatically a universal 95% guarantee. If uncertainty is too large to distinguish the intended acceptance limit, keep that decision unresolved or improve the reference and procedure. This guide does not issue a calibration certificate.
Distinguish hysteresis, creep and drift before refitting
Compare loading and unloading at matched reference values to inspect hysteresis. Under constant load, track deformation to inspect creep; under constant deformation, track force to inspect relaxation. Repeat an unloaded or known-reference check over the session to inspect drift. Keep loading rate, dwell time and previous contacts in the record so these effects are not collapsed into unexplained noise.
Costanzo and colleagues document material hysteresis and a nonlinear force–deformation relationship in their particular optical taxel design. This supports checking loading history; its numerical results cannot be assigned to DIGIT or GelSight Mini. A model trained only on increasing normal loads still needs independent evidence for unloading and tangential contact.
Common calibration failures
Keep acquisition faults separate from reconstruction and reference faults. Check the simplest observable layer first.
| Symptom | Inspect first | Useful next action |
|---|---|---|
| Baseline changes with no contact | Exposure, illumination, warm-up, mount and gel condition. | Retain before/after baselines; stabilize or record configuration changes before refitting. |
| Good central depth, poor edges | Spatial sampling, illumination and missing training locations. | Plot held-out residuals by surface location; expand reference coverage where justified. |
| Depth scale changes after resize | Crop, processing resolution and pixel-to-metric mapping. | Re-establish scale against the reference geometry; do not reuse an unrelated constant. |
| Marker jumps or missing tracks | Marker mask, contact deformation and tracking reinitialization. | Mark invalid spans; retain raw frames and avoid silently filling failed tracks with zero. |
| Different force on load and unload | Gel hysteresis, loading rate, reference alignment and time synchronization. | Record loading phase and independently validate the force model across the claimed conditions. |
| Low training error but poor new contacts | Leakage across adjacent frames, positions, objects or sessions. | Split by physical contact/session and test held-out geometries before reporting transfer. |
When should a calibration be reused, checked or refitted?
Treat a calibration as a versioned relationship between a particular sensor configuration and a stated operating envelope. A replacement pad is not automatically equivalent to the old pad, and a successful file load is not a transfer test. Use the following decision record after a maintenance or software change; it proposes a workflow rather than a universal recalibration schedule.
| Observed situation | Decision to evaluate | Evidence to retain |
|---|---|---|
| Configuration unchanged; reference checks meet predefined criteria | Reuse within the validated envelope, with planned monitoring. | Calibration ID, baseline and reference results, criteria and reviewer. |
| Offset changes, but sensitivity and geometry are not yet checked | Investigate; do not assume that zeroing restores the whole mapping. | Before/after zero, known-load or geometry checks and change history. |
| New gel, mount, crop, illumination or reconstruction revision | Validate transfer; refit if the mapping fails the intended criteria. | Exact configuration differences, held-out results and retained prior calibration. |
| Unstable reference, untraceable labels or untested load direction | Hold the affected measurement claim until the missing evidence is resolved. | Failure examples, unknowns and the next discriminating measurement. |
Write the release statement so another engineer can challenge it
Use a bounded statement: calibration ID; sensor and surface identity; target quantity and unit; tested area, load and environmental conditions; validation contacts and sessions; residual metrics and invalid rate; reference uncertainty; known exclusions; reviewer and date. Leave unknown fields blank or explicitly unknown. Do not substitute a broad label such as “robot ready” for those conditions.
Finally, evaluate the robot task separately. A geometry reconstruction can pass its reference checks while a grasp policy still fails. Keep calibration acceptance, controller performance and any safety assessment as separate records. The purpose is to make the next engineering decision defensible, not to turn a calibration error into a universal manipulation-success claim.
Connect calibration to processing and manipulation
Version each calibration and associate it with the observations that use it. Keep raw image or sensor units and validity flags alongside derived depth or force, recording the transformation and model revision. Sampling time, host receive time and end-to-end latency are distinct; a dataset timestamp does not substitute for those measurements.
For tactile sensing in robotics, the next question is whether the calibrated quantity supports a specific action or evaluation. A depth estimate may describe local contact geometry; a force estimate may support a separately validated contact task. Both need a task-specific protocol before a manipulation claim.
Common questions
Does 3D Cal provide a complete force calibration workflow?
The reviewed paper demonstrates RGB-to-gradient/depth reconstruction for DIGIT and GelSight Mini. Force-sensor integration is described as future work. Force estimation needs independent load references and a validated mapping.
Can I copy a calibration between two sensors?
Transfer is a hypothesis to test. Gel, camera settings, optics, mounting and wear can change the mapping. Validate against references on the receiving device before claiming accuracy.
What can I complete without hardware?
Read the source-based workflow, prepare the blank record template, and run the separate synthetic Python and dataset-checking exercises. Actual device calibration and error measurements require sensors and suitable reference equipment.
Next steps
- Robotics programming learning path →
Choose a data or tactile-feedback starting point.
- Process tactile CSV data with Python →
Generate heatmaps and teaching contact events without hardware.
- Compare tactile sensors →
Check measurement principles, interfaces and calibration requirements.
- Tactile manipulation →
Connect measurement quality to a contact task and its evaluation.
Source references
- 3D Cal: An Open-Source Software Library for Calibrating Tactile Sensors — checked 2026-09-16
- Costanzo et al. (2019): force/tactile sensor design and calibration — reviewed November 3, 2026
- NIST TN 1297: classification of uncertainty components — reviewed November 3, 2026
- NIST TN 1297: combined standard uncertainty — reviewed November 3, 2026
- NIST TN 1297: expanded uncertainty — reviewed November 3, 2026
- 3D Cal author project
- DIGIT official documentation — checked 2026-09-16
- DIGIT interface pinned source
- GelSight Mini official site — checked 2026-09-16
- GelSight gsrobotics pinned source