Results

Wear: 25 hours of gripping, no detectable effect

Within continuous segments — the only place a slope is attributable to use — only 3 of 7 beat a permutation null, and most deltas sit below their own floor. The drift does not accumulate across segments, so it is reversible per-session, not wear. LED warm-up is the natural candidate but this data cannot separate it from ambient temperature or gel creep, so no claim is made.

Does it hold on public data?

Finding 1 rests on a lab-internal campaign, so on its own it is a claim about one rig rather than about GelSight Minis. Two public datasets were checked, using their own reference frames and nothing from that rig.

Both reproduce the shape of the in-house result. Sparsh: 1 of 10 sessions drifts past the floor (6.6x), the other nine sit at 0.8–1.3x — the same heterogeneity as the 3-of-7 segments in the durability run. FoTa: the appearance steps at a 20-day gap in acquisition and then moves little across the twelve capture days that follow, and the step is roughly the size of swapping to the other finger on the same day (9.8 RMS). Neither dataset supports wear accumulating with use; both support appearance moving when the rig is disturbed.

Watch it happen

Drag the slider. The red ticks are breaks in acquisition, not in gripping time. The image jumps at a tick and then barely moves across the hours in between.

Markerless gel — 5.3 h. Most public force-labelled Mini data uses a markerless pad, so this is the run to read first.

Markered gel — 25.0 h, the longer run. Its 15 RMS step sits on the 88-hour break; the markerless run above instead shows a genuine rise inside its second segment. The two runs disagree on how much within-use drift there is, which is why neither is presented alone.

The reference frame is the adaptation

After changing a pad or sensor, recapture reference frames. It is unsupervised and recovers essentially the whole loss; skip it and physics-based force estimation stops working.

An earlier version of this page reported one condition that broke the pattern — calibration-free new→old surviving zero-shot at 0.93 — and flagged it as unexplained. It was an artefact. FEATS stores its frames as RGB and they were being read as BGR, so that comparison was measuring a channel swap rather than a sensor change. With the channels corrected it falls in line with the other three (0.10). n = 1 shift throughout.

Frozen encoders, compared at a common feature width

What are the encoders for?

Nothing here trains a network. Each encoder is a frozen feature extractor: a tactile image goes in, a fixed-length vector comes out, and only a small linear map from that vector to force is fitted. The question is not "which model is best at tactile sensing" — none of them has ever seen a tactile image in training. It is narrower and more useful: does a general-purpose visual representation survive a sensor change better than the physics pipeline does?

All four are ordinary computer-vision backbones, public and frozen, covering two pretraining regimes: ResNet50/18 from supervised ImageNet classification, and DINOv2 ViT-B/S from self-supervised training with no labels. They are the encoders someone would actually reach for. Two references anchor the comparison. The 32×32 raw-pixel control has no learned representation at all, so it marks the floor any encoder must beat. The T3 checkpoint is the one tactile-specific model available here.

Read the table this way: every row is fitted and scored the same way, so a difference between rows is a difference between representations, not between methods. In-domain rho sits in a narrow band across all six while transfer does not — so an in-domain number says nothing about robustness here. The T3 row is worst, but it is a fine-tuned checkpoint rebuilt from its tensors with an unstable width sweep; that is not evidence about the T3 method.

Does the shift predict the cost?

Recovering an unlabelled acquisition state from background colour

No dataset here labels force frames by sensor or pad. Background colour is set by the LEDs and the gel, not by what is pressed — so it can recover the missing state, but only measured on frames with no contact. Over the dataset's own 19 directories of genuine no-contact frames the within-directory spread falls to 0.017 against a cluster separation of 5.93 — a 340× margin.

Calibration-free reconstruction is unaffected; the per-sensor LUT degrades — the prediction this task existed to test. But the smaller cluster holds only two directories, the force histograms are not fully matched, and the LUT's recalibrated score being worse than its zero-shot score is physically implausible and remains unexplained.