RCT

Dataset & Benchmark · arXiv preprint · Under review at CoRL 2026

RCT: A Robot-Collected Touch–Vision–Language Dataset for Tactile Generalization

Jingbo He1,2, Michael Färber1, Roberto Calandra2

1TU Dresden & ScaDS.AI Dresden/Leipzig, Dresden, Germany  ·  2LASR Lab, TU Dresden, Dresden, Germany

TL;DRFrames from one robot press are near-duplicates. RCT preserves every press as a contact sequence, so tactile models can finally be evaluated on what deployment demands: unseen presses, positions, sensors — and unseen materials.

Overview of the RCT dataset: robot collection with three DIGIT sensors, contact sequences from shallow to deep indentation, the leakage problem in frame-random splits, and the ablation of overlap effects.
The RCT dataset preserves full contact sequences and enables held-out evaluation across materials, categories, sensors, contact positions, and robot presses.

At a glance

29,279tactile frames
122materials
7categories
1,832contact sequences
3DIGIT sensors
0.10 mmdepth step
15–17frames per press
|fz|force per frame
every frame indexed by material category sensor position sequence depth force

The dataset

Full robot presses on 122 industrial reference materials

A robot arm carries a rotating adapter that holds three DIGIT vision-based tactile sensors, so each material is recorded with all three sensors under one protocol. For every material, the robot presses the surface at multiple discrete contact positions; each press is sampled from initial contact to deeper indentation in fixed 0.10 mm steps, with contact force recorded at every frame.

The central design choice: tactile data is stored as contact sequences — the ordered frames from one press, on one material, with one sensor, at one contact position. Successive frames differ by a single 0.10 mm step, which makes them near-duplicate observations of the same physical interaction. Treating the press as the structural unit is what makes controlled held-out evaluation possible.

One contact sequence: DIGIT frames evolve gradually with indentation depth while measured force increases over the press.
One robot press. DIGIT frames evolve gradually with indentation depth while force increases — adjacent frames are correlated observations of one interaction.
Collection setup: robot arm with a rotating adapter holding three DIGIT tactile sensors above an industrial material sample.
Collection setup. A rotating adapter brings each of the three DIGIT sensors into the same contact pose over Musterkiste industrial reference materials.

Three aligned modalities

Touch

29,279 DIGIT frames in 1,832 contact sequences, each with indentation depth and per-frame contact force |fz|.

Vision

One 2048×1536 photograph per material — material identity, deliberately separated from contact state. Retrieval uses a material-level multi-positive criterion.

Language

Material-level descriptors from a fixed 54-term tactile vocabulary: two independent human annotators who physically touched every material (union is the default target), plus a VLM source (Gemma-3-27B) for scale.

Human–human descriptor agreement is moderate (Jaccard 0.294); VLM–human agreement is far lower (0.110 / 0.108) — visual descriptors capture only part of what people report from touch.

Material categories

Paper / Cardboard45
Plastic / Rubber33
Metal20
Textiles / Leather10
Wood / Bamboo / Cork7
Crafts4
Small Items3

122 industrial reference materials from the Musterkiste sample set (Modulor GmbH). Contact positions 1–4 are densely sampled across all materials; later positions are sparser and reported as metadata.


Why it matters

The independent unit is the press, not the frame

Within one press, the sensor advances 0.10 mm at a time — so a frame-random split routinely places near-duplicate observations of the same physical interaction in both training and test. A model can then retrieve a near-duplicate contact instead of learning a representation that transfers. Try it below.

Interactive · one real robot press · material_0701420

Photograph of material_0701420 — a perforated grid panel — the same material being pressed below.

A real press from the dataset, the same material seen by two modalities: the vision photo of material_0701420 (a perforated grid panel), and below, the live touch frames as a DIGIT sensor presses into it — the grid imprints into the gel.

Position Sensor
material_0701420 rest + 16 frames · Δ0.10 mm
Live DIGIT tactile frame from material_0701420: starts at the no-contact rest frame, then the perforated-grid texture imprints into the gel as the press deepens.

rest frame · no contact · sensor 1

restindentation
|f z| · force log
1 / 17frame
no contact (rest)Δ = 0.10 mm / frame1.5 mm

Drag the slider — or click any frame below. It starts at the no-contact rest frame, then steps through the 16 real DIGIT contact frames of one press on material_0701420; once in contact, adjacent frames barely change. That is the point: frames within one contact sequence are near-duplicates, and any split that separates them frame-by-frame leaks the interaction itself.

Now split this press for evaluation

Decomposing frame-random performance

80.0
−17.7 pp →
62.3
−42.0 pp →
20.4
Frame-random
(full overlap)
No contact-sequence
overlap
No sequence &
no material overlap

Tactile→text Recall@1 on RCT, one representative held-out split (seed 42). The first drop is measured with the encoder held fixed — only the test pool changes — and an encoder trained directly under the held-out-sequence protocol lands at 62.44%, confirming the gap is the protocol, not a weaker representation. Across three held-out draws, held-out-material R@1 averages 25.1 ± 6.1%.

The released TVL/HCT split shows the same structure

Auditing the public split with its own files: every HCT test contact sequence also appears in training, and the overlap is exploitable without any learning at all.

279 / 279test contact sequences
also present in training
1 stepmedian gap to the nearest
same-sequence training frame
98.3%top-1 sequence recovery by a
training-free raw-pixel nearest neighbor

This is an evaluation-protocol result, not a criticism of touch–vision–language learning — RCT uses the same training recipe. It means frame-random retrieval scores should not be read as tactile generalization unless contact-sequence and material overlap are controlled.

A held-out sensor is as hard as a held-out material

Held-out axes · RCT-only training · tactile→vision uses the multi-positive criterion
Held-out axis (materials seen at train)Tac→text R@1Tac→vis R@1
Contact position (p5; seen at p1–p4)65.46%74.16%
DIGIT sensor instance (d3; seen at d1, d2)21.51%13.00%
Held-out material (K=20; not seen)20.35%18.87%

Holding out a contact position is easy — the material is still seen elsewhere. Holding out one DIGIT instance is as hard as holding out the material entirely: sensor-instance robustness is its own bottleneck, and honest evaluation needs sensor-disjoint splits.

Uniformly sampling the press beats using every frame

Training-time frame sampling · same held-out-contact-sequence test pool
Training densityFramesTac→text R@1Tac→vis R@1
Full press (~16 frames / sequence)22,57662.44%70.95%
Deep5 (five deepest frames)7,05559.52%65.96%
Uniform5 (five frames, shallow→deep)7,05568.18%78.29%

Covering the press from shallow to deep contact (+5.7 pp over full density, at one third of the data) is more useful than densely sampling near-duplicates — consistent with near-duplicates acting as false negatives in contrastive objectives.

RCT training improves material structure on unseen materials

Frozen embeddings on K=20 held-out materials
EncoderSeparability marginCategory probe (7-class)Hard/soft probe
Released TVL encoder0.04834.2%52.2%
Random init0.00437.8%45.7%
RCT encoder0.29749.8%52.2%

Category-level material information transfers to unseen materials; binary hardness stays at or below the majority baseline (59%) for every representation — including force features — marking it as an open problem.


Benchmark protocols

Hold out physically meaningful units

RCT ships split files and a split-generation toolkit for six evaluation settings. Held-out material (K = 20) is the primary protocol, because it matches the deployment question: can a robot recognize material properties it has never touched?

Held-out settings supported by RCT
SettingHeld out at test timePurpose
Held-out materialAll sequences of K materials (K = 2, 5, 20)Unseen materials
Held-out categoryAll materials from one top-level categoryUnseen material family
Held-out sequenceEntire contact sequences; materials may be seenSequence overlap
Held-out positionOne contact position across materialsNovel location
Held-out sensorOne DIGIT sensor instanceSensor transfer
Frame-random controlRandom frames; material and sequence overlapSplit inflation

Reporting recommendations

Tactile benchmarks should expose contact-sequence identifiers and report held-out results alongside any frame-random score, stating the level at which test samples are independent of training. For vision-based tactile sensors that level is at least the contact sequence; for manipulation it is usually the material. Because a single K = 20 draw is noisy (±6.1 pp across seeds with diverse selection; 29.8 ± 2.9% with category-balanced selection), report the mean and standard deviation over multiple held-out draws rather than a single split.


Download

Get the dataset

The release contains everything needed to reproduce the paper's protocols and to build new ones:

  • 29,279 tactile frames with contact-sequence identifiers
  • Per-frame force signals and indentation depths
  • One 2048×1536 photograph per material
  • Material-level descriptor annotations (two human annotators, human-union, and VLM source) over a 54-term vocabulary
  • Material / category / sensor / position metadata
  • Split files for all held-out settings + split-generation and evaluation scripts

License: CC BY 4.0. One material had an ID-format mismatch in early annotator labels; the released metadata already includes the correction.

Hosted on Figshare

Download RCT

figshare.com/s/a5ed417ba6602ccad0f6


Citation

BibTeX

arXiv preprint entry — fill in the ID and author list after the announcement; switch to the CoRL/PMLR entry on acceptance.

@misc{he2026rct,
  title         = {{RCT}: A Robot-Collected Touch--Vision--Language
                   Dataset for Tactile Generalization},
  author        = {Jingbo He and Michael F{\"a}rber and Roberto Calandra},
  year          = {2026},
  eprint        = {2606.31694},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2606.31694},
  url           = {https://arxiv.org/abs/2606.31694},
}