# World-model dataset audit: unmarked GelSight Mini

Audit date: 2026-09-20. Worktree: `feature/tactile-activity-benchmark`.
Scope: OmniViTac/OmniVTA and ContactWorld only. Inclusion requires affirmative
evidence of **real GelSight Mini without physical markers**, an identifiable
released subset, and usable original tactile image sequences. A marker-removed
or background-subtracted image does not meet this requirement.

## Inclusion decision

| Candidate | Decision for React + ToucHD + MiTaS benchmark | Positive evidence and remaining uncertainty |
|---|---|---|
| OmniViTac / OmniVTA | **Do not include. No unmarked Mini subset verified.** | Current paper Fig. 11 photographs a Mini with a regular physical dot grid. A selectively inspected released episode has 700-point deformation at approximately 60 Hz and 700×400 difference images, matching the documented Xense setup, not Mini. The bulk release lacks a sensor-level manifest. This does **not** prove every unindexed trajectory is marked or that no unmarked Mini recording exists. |
| ContactWorld | **Do not include. Observed real Mini data are marked; real training release not located.** | Current paper Fig. 11 shows a dot grid in TacRGB for all four real tasks; §V explicitly describes marker tracking. The identifiable public dataset releases are the 12 simulation tasks. No released unmarked real Mini episode was verified. |

No benchmark manifest or cohort was changed. These are exclusions from the
current evidence-qualified cohort, not universal claims about unpublished or
unindexed data.

## Publication selection

The selected publication consists of **this report and the two sensor crops only**:
`omni_fig11_sensor_crop.png` (77,810 bytes) and
`contact_fig11_tactile_crop.png` (586,358 bytes). Together with the report this is
less than 0.7 MB. Preserve the relative evidence paths when copying the crops.
Primary paper, project, repository and pinned dataset links below provide the
public source trail. No commit or upload was performed by this audit.

All archive ranges, pickle prefixes/samples, original downloaded figures,
project videos and other binary evidence remain local and are excluded from
this publication selection. The evidence directory is a local audit inventory,
not a directory to publish recursively.

Small source JSONs retained locally for optional reviewer handoff are
`figure_sources.json` (869 bytes), `sources_manifest.json` (26,626 bytes),
`release_summary.json` (2,402 bytes), and `download_budget_summary.json`
(280 bytes), plus `verification.json`. They are **not** part of the requested
three-file publication. `publish_selection.json` records the exact selected
paths, hashes and source URLs; the complete local inventory is separate.

## Evidence directory

All supporting files are under
`artifacts/tactile_activity/worldmodel_mini_audit/`（完整证据保存在本地审计目录）.
The directory contains source snapshots, complete remote inventories, bounded
archive ranges, literal-pickle metadata inspections, original paper figures,
project demonstration videos, extracted frames, and SHA-256 manifests.

The clearest visual evidence is:

- [OmniVTA Fig. 11, sensor crop](audit/omni_fig11_sensor_crop.png): regular dots on the photographed Mini gel. [Original complete figure](https://arxiv.org/html/2603.19201v3/tsne.png).
- [ContactWorld Fig. 11, tactile row](audit/contact_fig11_tactile_crop.png): physical dot grids in all four TacRGB panels. [Original complete figure](https://arxiv.org/html/2606.13877v2/real.png).
- [OmniViTac project demo at 6.94 seconds](https://mrsecant.github.io/OmniVTA/assets/moive/dataset.mp4): separate columns explicitly labeled `tactile raw`, `tactile diff.`, and `deformation`. The raw panels have dense dots; the difference panels suppress their appearance. This mixed-sensor demonstration is **not independently identified as Mini**.

The crops only select pixels from the original figures. They do not remove dots,
apply enhancement, or synthesize content.

## OmniViTac / OmniVTA

Primary links:

- [Paper v3, 2026-08-10](https://arxiv.org/html/2603.19201v3)
- [Project](https://mrsecant.github.io/OmniVTA/)
- [Model repository](https://github.com/MrSecant/OmniVTA)
- [Dataset at audited revision](https://huggingface.co/datasets/tars-robotics/OmniVitac/tree/060e506ca3cc49571e7ba4db4fc97e24d7711c0d)
- [Original Fig. 11 image](https://arxiv.org/html/2603.19201v3/tsne.png)

### Sensor and paper evidence

The paper reports 21,879 trajectories, 86 tasks and 126 objects **across four
sensor families**, predominantly Xense. Mini is specified as 320×240 RGB plus
9×7 displacement at 25 Hz. Xense is specified with 700×400 tactile imagery and
35×20 displacement, the latter at 60 Hz. Fig. 11 provides direct photographic
evidence of a marked Mini used in the tactile representation experiments.
The paper describes timestamped acquisition and aligned visual, tactile and
action streams, but these aggregate descriptions do not establish an unmarked
Mini release or its episode count.

### Actual release inventory

HF revision: `060e506ca3cc49571e7ba4db4fc97e24d7711c0d`.
The complete recursive listing contains 738 files totaling
**38,616,022,846,761 bytes**. Of these, 736 are archive parts:

| Category directory | Parts |
|---|---:|
| `Adjustment_zip_new` | 130 |
| `Assembly_zip_new` | 134 |
| `Cutting_zip_new` | 67 |
| `Grasping_zip_new` | 300 |
| `Peeling_zip_new` | 53 |
| `Wiping_zip_new` | 52 |

The only small top-level files are `.gitattributes` and a 33-byte, license-only
README. There is no visible episode/sensor index or Mini-specific directory.
Most parts are 52,613,349,376 bytes each. `Peeling.tar.gz.03` is absent.
The [open extraction issue](https://huggingface.co/datasets/tars-robotics/OmniVitac/discussions/2)
reports the same missing part; its absence was independently checked against
the saved inventory. Archive availability is not proof of complete extractability.

Saved evidence: `omni_hf_info.json`, `omni_hf_tree.json`,
`release_summary.json`, and `omni_github_tree.json`. The model repository tree
contains LICENSE only; it provides no loader or sensor mapping usable for this
audit. The advertised hardware repository returned 404 in the preceding audit.

### Bounded inspection of released contents

The first archive part of **each of the six categories** was opened with HTTP
Range and streaming gzip/tar traversal. Every sampled beginning placed
`depth.pkl` before tactile data. The first depth payloads ranged from
736,083,764 to 1,314,844,813 uncompressed bytes. These are file sizes, not downloads.
Logs: `omni_*_prefix.gz.probe.json`.

One concrete episode was followed farther:

```text
Remote: Assembly_zip_new/Assembly.tar.gz.00
Archive prefix:
mnt/oss_data/it20260119/Assembly/
  Assembly_Gello_Insert the USB drive_100/0000/
```

[Pinned remote part](https://huggingface.co/datasets/tars-robotics/OmniVitac/resolve/060e506ca3cc49571e7ba4db4fc97e24d7711c0d/Assembly_zip_new/Assembly.tar.gz.00)

| Member | Uncompressed bytes | What was actually inspected |
|---|---:|---|
| `depth.pkl` | 736,083,764 | Header; traversed/discarded, not loaded |
| `gripper.pkl` | 22,223 | Complete literal opcode inspection; `gripper_pos`, `timestamps` |
| `image.pkl` | 539,168,978 | Prefix: `camera1/image`, first array 480×640×3 `uint8`; not tactile |
| `state.pkl` | 114,946 | Complete opcode inspection; `eef_pose`, `joints`, `timestamps`, `joint_stamps` |
| `tactile.pkl` | 40,801,831 | Complete opcode inspection; two tactile groups, 700×3 `float64` deformation arrays, timestamps |
| `tactile_diff.pkl` | 1,084,512,496 | 2 MiB prefix; `tactile1/diff_image`, 700×400×3 `uint8` arrays |

The inspected deformation shape is 700 points (=35×20), **not** Mini's 63
(=9×7). Together with the image shape and cadence, this is strong evidence for
Xense; the files themselves do not contain an explicit manufacturer string.
Accordingly this is a sensor inference, not a claimed Mini sample.

Literal timestamp results, assuming milliseconds from their epoch magnitude:

| Stream | Samples | First / last timestamp | Rate from span |
|---|---:|---|---:|
| Gripper | 1,304 | 1769681670737 / 1769681692453 | 60.0018 Hz |
| EEF state | 1,304 | 1769681670734 / 1769681692451 | 59.9991 Hz |
| Tactile deformation 1 | 1,210 | 1769681670741 / 1769681690881 | 60.0298 Hz |
| Tactile deformation 2 | 1,212 | 1769681670741 / 1769681690914 | 60.0307 Hz |

No nonpositive timestamp increments occurred within these streams. Their
different counts/end times must not be treated as already row-aligned.
These rates concern the sampled Xense-like **deformation/state** streams;
neither a Mini rate nor the tactile RGB acquisition rate was measured.
Recorded EEF states are not independently verified commanded-action targets.

`assembly_state_timestamp_audit.json` and `inspect_sample_metadata.py` reproduce
these measurements without executing pickle globals. Two difference-image
arrays were recovered from literal byte opcodes and saved with both stored and
R/B-swapped channels, since channel order was not declared. They are explicitly
**processed difference images**, not qualifying raw unmarked images.

Traversal stopped inside `tactile_diff.pkl` at the chosen range limit. No
original tactile RGB member was reached. It would be incorrect to infer from
that stopping point that the archive lacks raw tactile RGB. It would likewise
be incorrect to attribute this Xense-like sample to Mini.

### Demonstration media

The [project dataset video](https://mrsecant.github.io/OmniVTA/assets/moive/dataset.mp4)
is downloadable: 38,860,935 bytes, 46.278 seconds, encoded at 30 fps. Its raw
tactile panels visibly contain dots; its processed difference panels can look
dot-free. This edited montage does not provide native per-sensor timestamps or
a released Mini episode identity. See `media_ffprobe.json` and
`omni_dataset_demo_{0.15,0.5,0.85}.png`.

## ContactWorld

Primary links:

- [Paper v2, 2026-09-08](https://arxiv.org/html/2606.13877v2)
- [Project](https://contact-world.github.io/)
- [Code](https://github.com/PokuangZhou/ContactWorld)
- [Official dataset at audited revision](https://huggingface.co/datasets/Pokuang/ContactWorld/tree/6587189d8e06480e739af5a4004d411c5c07743b)
- [Author's OXT-format release](https://huggingface.co/datasets/ZhiyuanZhangZhiyuan/ContactWorld/tree/f7f3ea533262eeb6c28bacc9d45e5fcc6ee761dd)
- [Original real-task Fig. 11](https://arxiv.org/html/2606.13877v2/real.png)

### Current real Mini evidence

Version 2 describes Mini for four physical tasks: USB insertion, elastic-clip
removal, valve rotation and ball sorting. It reports 80 demonstrations per
task (50 training, 30 held out), totaling 320 demonstrations, and separately
360 physical evaluation rollouts. All four Fig. 11 TacRGB panels visibly show
a regular marker grid. Section V-A also states that TacFF combines surface
deformation with tracked marker displacements.

The v1 description of R1.5 and 65 valve demonstrations is superseded for the
current real experiment; it must not be used to dismiss v2's Mini identity.
Conversely, v2's Mini identity must not be assigned to older simulated releases.
The explicit 10-Hz acquisition/control description accompanies simulation;
native real Mini RGB acquisition fps and original timestamps were not verified.

### What is actually downloadable

Official HF revision: `6587189d8e06480e739af5a4004d411c5c07743b`.
The inventory contains 12 dataset tarballs and 12 checkpoint tarballs under
`releases/`, plus assets/metadata: **133,840,587,165 bytes** overall.
Task names match the 12 simulation variants. No separately named real Mini
release appears in this complete listing.

A 1-MiB HTTP range from
[`releases/insertion_usb_dataset.tar.gz`](https://huggingface.co/datasets/Pokuang/ContactWorld/resolve/6587189d8e06480e739af5a4004d411c5c07743b/releases/insertion_usb_dataset.tar.gz)
(full size 6,585,153,409 bytes) exposes Zarr arrays `data/dof_vel` and
`data/dof_pos`, each shaped **[13671, 9]**. The 13,671 rows match the paper's
simulation USB task, providing file-level corroboration rather than relying
only on the archive name. See `contact_usb_prefix_fullscan.json`.

The second author-hosted release, revision
`f7f3ea533262eeb6c28bacc9d45e5fcc6ee761dd`, contains 12 Zarr ZIP archives and
metadata totaling **63,035,445,386 bytes**. Its README explicitly identifies
the data as simulated; the tactile note declares `GelSightWedgeR1.5`, not Mini.
It documents RGB, depth, and 10×14×3 TacFF packages, timestamps and episode
boundaries. That format does not turn simulated data into a real Mini corpus.
See `contact_oxt_readme.md`, `contact_oxt_tactile_note.json`, and
`contact_oxt_tree.json`.

An HF search for ContactWorld returned these two repositories. The pinned
GitHub tree (`8d0d0edff44f2bb28b58d96e3cf638088589e789`) also exposes no real
episode files; `data/demo_data` contains `.gitkeep`. No claim is made about
unlinked/private releases outside this search.

### Real-world videos are demonstrations, not raw tactile release

Two current project videos were downloaded and inspected:

| Public file | Bytes | Duration | Encoded average fps |
|---|---:|---:|---:|
| [real_usb.mp4](https://contact-world.github.io/videos/real_usb.mp4) | 7,991,368 | 11.893424 s | 10193616/347999 (~29.292) |
| [real_valve.mp4](https://contact-world.github.io/videos/real_valve.mp4) | 3,221,705 | 8.098413 s | 3444184/117997 (~29.189) |

These are edited multi-panel comparisons of robot executions. They do not
supply original tactile RGB streams with capture timestamps or action arrays.
The encoded video rates must not be reported as Mini acquisition fps.

## Bounds, reproducibility and verification

- Started at 06:36 UTC; investigation and reporting stayed within the authorized
  15-minute window.
- Download plan: small current source/inventory snapshots and figures; three
  bounded project videos; six initial Omni archive ranges; one ContactWorld
  range; then sequential extensions of one Omni episode with explicit byte and
  traversal-time caps. Every archive request required HTTP 206.
- Saved archive-range bodies plus project videos total **674,505,112 bytes**.
  HTML, JSON and figure downloads add approximately 10 MB. Total remains below
  the approximately 1-GB sample allowance. Extracted copies are not additional
  network downloads.
- All large gzip artifacts are deliberately incomplete prefixes or continuation
  segments. Expected end-of-data outcomes are scope limits, not evidence of
  corrupt upstream files. They should not be passed to a full-dataset loader.
- No raw pickle was executed. `pickletools.genops` inspected literal strings,
  numbers, shapes and byte payloads; NumPy interpreted only explicitly sized
  literal byte arrays for difference-image previews.
- Source and artifact manifests retain URLs, pinned revisions, sizes, range
  boundaries and SHA-256 hashes. Local linked evidence was checked for existence;
  JSON snapshots and image files were parsed/verified; the report was checked
  for whitespace errors. No pipeline code changed, so no pipeline test suite was
  run. Nothing was committed or published.

Reconsider inclusion only after locating a **specific original unmarked Mini
episode**, with sensor/gel provenance, raw RGB frames, timestamps or defensible
native cadence, and an exact downloadable subset path. Neither candidate
currently satisfies that gate.