Hugging Face's logo

Datasets:
BrainCoTech
/
RevoTac-OXT

Languages:
English
Size:
n<1K
Tags:
robotics
tactile
simulation
zarr
open-x-tactile
Dataset card Files
xet
 Community
RevoTac-OXT / README.md
DavidLXu's picture
DavidLXu
Group all task Zarr archives under the RevoTac-OXT dataset directory
0104b9b verified
preview code
|
Raw History Blame Contribute Delete
7.46 kB
metadata 
pretty_name: RevoTac OXT — Ten Simulated Tactile Manipulation Episodes
language:
  - en
tags:
  - robotics
  - tactile
  - simulation
  - zarr
  - open-x-tactile
size_categories:
  - n<1K

RevoTac OXT

Ten scripted simulated tactile manipulation episodes, one per task, from RevoBench / RevoSim. Contributing institutions: HKUST(GZ), BrainCo, Georgia Tech. Contact: Lixin Xu, davidxulixin@gmail.com.

Project: https://braincotech.github.io/RevoSim-web/
Code: https://github.com/BrainCoTech/RevoBench
Format and contribution requirements: Open-X-Tactile and OXT-QC.

This release follows section 4.1 of the OXT submission guide: the single dataset directory RevoTac-OXT/ contains all ten <task>.zarr.zip archives. No task archive or unpacked Zarr is stored at the repository root. Uploading this repository does not mean that OXT organizers have accepted the dataset.

File layout 

BrainCoTech/RevoTac-OXT
├── RevoTac-OXT/
│   ├── cup_hang.zarr.zip
│   ├── hammer.zarr.zip
│   ├── insert_peg.zarr.zip
│   ├── insert_pipette.zarr.zip
│   ├── kitchen.zarr.zip
│   ├── plug_charger.zarr.zip
│   ├── pour_can.zarr.zip
│   ├── power_drill.zarr.zip
│   ├── rotate_knob.zarr.zip
│   └── tray_carry.zarr.zip
├── README.md
├── metadata.json
├── manifest.json
├── citation.bib
└── qc/

The repository, dataset directory and current metadata.name all use RevoTac-OXT. Earlier audit reports retain the historical source identifier RevoBenchSimTactile10NoIdle. The directory update only changes repository paths and descriptive metadata; the ten archive SHA-256 hashes and all Zarr values are unchanged.

Contents

Version 0.1.0 contains 10 episodes, 32,272 frames at 60 Hz. The unpacked Zarr files total 62,347,719,867 bytes (62.348 GB). Exact archive sizes, SHA-256 hashes and source file counts are in manifest.json.

Task Frames Archive
cup_hang 4178 RevoTac-OXT/cup_hang.zarr.zip
hammer 1856 RevoTac-OXT/hammer.zarr.zip
insert_peg 3406 RevoTac-OXT/insert_peg.zarr.zip
insert_pipette 2937 RevoTac-OXT/insert_pipette.zarr.zip
kitchen 5655 RevoTac-OXT/kitchen.zarr.zip
plug_charger 2671 RevoTac-OXT/plug_charger.zarr.zip
pour_can 4653 RevoTac-OXT/pour_can.zarr.zip
power_drill 2866 RevoTac-OXT/power_drill.zarr.zip
rotate_knob 2459 RevoTac-OXT/rotate_knob.zarr.zip
tray_carry 1591 RevoTac-OXT/tray_carry.zarr.zip

Each ZIP contains one <task>.zarr/ directory, including data, auxiliary observations and edit_provenance. Extract all entries; do not discard auxiliary or edit-provenance arrays. ZIP64 is used. ZIP storage preserves the already-compressed Zarr chunk bytes.

Download and open 

from huggingface_hub import hf_hub_download
from pathlib import Path
import hashlib
import json
import zipfile
import zarr

repo = "BrainCoTech/RevoTac-OXT"
manifest = json.loads(Path(hf_hub_download(repo, "manifest.json", repo_type="dataset")).read_text())
entry = next(x for x in manifest["archives"] if x["task"] == "cup_hang")
archive = Path(hf_hub_download(repo, entry["path"], repo_type="dataset"))
digest = hashlib.sha256()
with archive.open("rb") as stream:
    for block in iter(lambda: stream.read(16 * 1024 * 1024), b""):
        digest.update(block)
assert digest.hexdigest() == entry["sha256"]
with zipfile.ZipFile(archive) as packed:
    packed.extractall("data")
episode = zarr.open_group("data/cup_hang.zarr", mode="r")
print(episode.tree())

Signals and editing

The bimanual robot has 56 joint coordinates. The dataset includes measured joints and commands, four scene RGB cameras, canonical ego/wrist poses, five optical tactile pads per hand, numeric indentation depth in meters, marker displacement, pad-contact wrenches, and 285 pressure taxels per hand. Stiffness and damping are omitted from this OXT view.

Pressure is a nonnegative normalized response, not Pa or N. DIP optical pads and palm/MCP/PIP pressure regions have different spatial coverage. The sensor facade is target-specific. Simulated Taxim RGB marker warping is not implemented; these are not real sensor recordings.

Long stationary holds were removed offline, reducing 34,528 original frames to 32,272. Modalities share the same retained source-frame indices. Native retained values, sensor resolution and chunk codecs are preserved; timestamps were reindexed at 60 Hz. Original capture frame indices/times are retained in edit_provenance. No interpolation or physics rerun was performed. These ten development demonstrations are not a statistical task-success benchmark.

Quality checks and known limitations

See qc/ for the review records. The official metadata check passed all ten dataset structures; its historical report still notes the then-missing download URL and FTP-1 code_trigger. This release fills metadata.json.url; code_trigger remains unassigned. The release metadata check was rerun after adding the URL: all ten dataset structures passed, with only code_trigger still missing (see qc/release_metadata_check/).

The official external VLM stage was attempted without dry-run but could not produce scores because no API key was available. Its reports are preserved with missing_api_key and null scores. A separate Codex multimodal agent review examined sampled scene/both-hand tactile evidence plus all-frame numerical checks. This does not constitute organizer approval or a completed external VLM API check. Published JSON reports replace internal source-directory prefixes with descriptive placeholders; the original private audit reports remain unchanged.

  • pour_can: right pinky force reaches 416.2466 N at retained/original frame 600, an isolated peak much larger than neighboring values. The value is preserved and flagged for contact-model review.
  • plug_charger: the left thumb DIP records contact in only eight frames. Source validation supports successful insertion and post-release stability, but does not establish continuous three-fingertip contact.
  • rotate_knob: peak angle is about 34.16 degrees; the knob returns near zero after release. This is not a demonstration of maintaining a 35-degree final angle.
  • Active-hand pressure is zero throughout eight tasks. Optical and pressure regions differ; these pressure channels should not be described as rich contact observations.
  • 4D recordings use a full-duration 10 Hz inspection view with 320x240 RGB and documented nominal intrinsics/display adaptations. No scene depth was available, so point clouds were disabled. All 32,272 native pose frames were checked numerically against previously validated FK sources. The recording is sampled coordinate-review evidence, not a full native-rate task video or pixel-accurate calibration certificate.

The kitchen task targets one potato piece and two meat pieces; carrots remaining in the pan are distractors. The drill task requires alignment above the board, not actual drilling. The pipette task ends with the grasp maintained.

Citation

See citation.bib for the citation supplied on the project website. No paper identifier is inferred. A separate dataset license has not been specified in this release.