Original data audit / 2026 edition

Tactile robotics dataset transparency audit

A reproducible audit of what 19 source-reviewed tactile robotics dataset records disclose about public data URLs, direct license links, code, sampling rates, synchronization, and data splits.

Published and reviewed 2026-08-27 by . Source records were last reviewed through 2026-08-22.

Reusable data

The human-readable table, CSV, and JSON are generated from the same records and the same deterministic checks.

Computed findings

Disclosure is measurable only when the rule stays narrow

These counts are generated from the current directory. They do not turn a URL, keyword, or repository into a quality score. Each card states exactly what passed the audit rule.

Dedicated dataset URL
7/1937%

A dataset or file-hosting URL is present in the reviewed record.

Direct license URL
7/1937%

A machine-verifiable license URL is present, not only a prose license name.

Code repository
11/1958%

A public GitHub repository URL is attached to the record.

Sampling rate mentioned
3/1916%

The audited record text contains an explicit numeric Hz or FPS expression.

Synchronization mentioned
10/1953%

The audited record text mentions synchronization or timestamps.

Train/test/split mentioned
7/1937%

The audited record text mentions a training, validation, test, or split signal.

Record-level evidence

Every yes and no is inspectable

“No” means the field or disclosure signal did not pass this audit rule in the current RoboSkin.ai record. It does not prove that the information is absent from every version of the underlying project.

Record-level disclosure checks for every tactile robotics dataset entry in this audit.
Dataset recordYearData URLLicense URLCodeRateSyncSplitReviewedPrimary source
UniVTAC Encoder Pretraining Corpus2026NoNoYesNoYesYes2026-08-22Paper ↗
UniVTAC Benchmark Dataset2026YesYesYesNoNoYes2026-08-22Paper ↗
T-Rex Tactile-Reactive Dexterous Manipulation Dataset2026YesYesYesYesNoNo2026-08-22Paper ↗
EgoTouch2026YesNoYesYesYesYes2026-08-22Paper ↗
PRISM: Precision and contact-rich Real-world Industrial Skill dataset with Multimodal sensing2026NoNoYesNoYesNo2026-08-22Paper ↗
SoftVTBench2026YesYesYesNoYesNo2026-08-22Paper ↗
RoboTacDex: A Dexterous Visual-Tactile-Action Dataset for Humanoid Manipulation2026NoNoNoYesYesYes2026-08-22Paper ↗
HT-Bench2026NoNoNoNoYesYes2026-08-22Paper ↗
RCT: Robotic Contact Tactile2026YesYesYesNoNoYes2026-08-16Paper ↗
TactiDex2026NoNoNoNoYesNo2026-08-16Paper ↗
FreeTacMan2025YesYesYesNoNoNo2026-08-16Paper ↗
Humanoid Visual-Tactile-Action Dataset2025NoNoNoNoYesNo2026-08-16Paper ↗
Sparsh-X Multisensory Touch Resource2025NoNoNoNoYesYes2026-08-16Paper ↗
Touch and Go2022NoNoNoNoYesNo2026-08-19Paper ↗
TVL: Touch, Vision, and Language2024NoNoNoNoNoNo2026-08-19Paper ↗
ObjectFolder Real2023NoNoYesNoNoNo2026-08-19Paper ↗
ObjectFolder 2.02022YesYesYesNoNoNo2026-08-19Paper ↗
TacVerse2026NoNoNoNoNoNo2026-08-19Paper ↗
VTDexManip2025NoYesYesNoNoNo2026-08-19Paper ↗

Method

A bounded, reproducible text audit

The audit covers every record in the RoboSkin.ai tactile robotics dataset directory on the publication date. It is a bounded editorial sample, not a census of every tactile dataset in existence.

  1. 01Dataset URL means the reviewed record contains a dedicated dataset or file-hosting URL. It does not prove that every file, revision, or advertised subset is downloadable.
  2. 02License URL means the record contains a direct URL for the dataset license. A license named only in prose does not pass this stricter machine-verifiable test.
  3. 03Code repository means the record contains a public GitHub repository URL. It does not establish that the dataset itself or every model weight is licensed under the repository license.
  4. 04Sampling rate, synchronization, and split signals are deterministic keyword checks over the record’s sample-count, format, availability, modality, and sensor fields. They measure documented disclosure in the reviewed record, not independent validation of the underlying files.

Limitations

What this audit cannot establish

  • RoboSkin.ai did not download and checksum every hosted package for this audit.
  • A missing signal can mean either that the primary public material did not state it clearly or that the current editorial record has not captured it yet.
  • A present signal does not establish data quality, calibration quality, benchmark validity, or suitability for a particular robot.
  • Records differ in scope: some describe released trajectory packages, while others describe research corpora that were not verified as standalone downloads.

Corrections should identify the record, the exact field, and a primary source. The audit will change only when the underlying structured record changes.

Submit a source or correction ->