Datasets:
image imagewidth (px) 1.79k 1.79k | label class label 0
classes |
|---|---|
null | |
null | |
null | |
null | |
null |
3dtacdex
This dataset follows the Open-X-Tactile (OXT) submission format and FTP-1 data-processing conventions. The OXT checklist and organizer package document completed preparation and remaining scientific limitations; official acceptance is pending.
Institution: 北京大学(Peking University)
Contact: kingchou007@gmail.com
Snapshot: local-oxt-task-merged-v2 (local conversion identifier; final release version and license are not yet specified).
Processed tactile manipulation data for a right LEAP hand on a JAKA arm: 47,124 unique frames, 328 continuous segments, 4 task stores, two fixed RGB views and eight PaXini tactile pads. Same-task versions are concatenated into one store per task: assembly, flip, play_data, openbox.
The 11 retained source/version groups remain traceable through data/source_group and root lineage metadata. data/source_split preserves play train/test identity. Episode boundaries are retained in meta/episode_ends; concatenation does not join independent trajectories or reset their original timestamps.
All four retained stores pass OXT structure checks with zero dataset errors or warnings. Actual local VLM screening is completed with warnings: 20 standard pages across all four retained tasks (4 task warnings, 0 failures), plus 11 diagnostic sheets. The owner confirmed that the four all-zero middle/ring tactile pads in openbox are expected. Model misreadings and sampling limits are documented in the VLM review. This public review draft has not received formal OXT approval. See current status, current scope audit, and quality report.
Download and read
The distributed files are 4 lossless ZIP64 archives, one Zarr v2 store per task. Existing Zarr compression is preserved (ZIP_STORED); archives total approximately 174 GB. The original four source ZIPs, raw staging copies and model weights are retained locally and are not uploaded. Each <group>/manifest.json records the archive SHA256 and every member's size, CRC and SHA256. checksums.sha256 covers the distributed archives.
hf download kingchou007/3dtacdex --repo-type dataset --local-dir 3dtacdex
cd 3dtacdex
sha256sum -c checksums.sha256
python -m pip install -r requirements-qa.txt
python verify_dataset.py
Newer hf CLI releases may use --type dataset in place of --repo-type dataset; check hf download --help. The QA script reads directly from ZIP files. A notebook runs the same checks. To read only one group without extracting:
from pathlib import Path
import zarr
p = Path('openbox/openbox.zarr.zip')
with zarr.ZipStore(str(p), mode='r') as store:
root = zarr.open_group(store=store, path=p.name[:-4], mode='r')
rgb = root['data/camera_main_rgb'][0]
tactile = root['data/right_tactile_data_fingers'][0] # (8, 15, 3)
episode_ends = root['meta/episode_ends'][:]
For directory-based loaders, extract each archive into zarr/; it already contains the <group>.zarr/ root. Allow approximately another 174 GB of disk space for a full extraction. Use Zarr 2, and open stores read-only.
Contents and schema
| Task | Frames | Continuous segments | Source/version groups |
|---|---|---|---|
| assembly | 22,069 | 124 | 4 |
| flip | 12,570 | 80 | 4 |
| openbox | 3,671 | 50 | 1 |
| play_data | 8,814 | 74 | 2 |
All canonical arrays are under data/ with first dimension T. meta/episode_ends stores exclusive segment boundaries.
| Field | Shape / convention |
|---|---|
camera_main_rgb |
(T,720,1280,3), uint8 RGB |
camera_secondary_rgb |
(T,768,1024,3), uint8 RGB, additional fixed external camera |
timestamps |
Source timestamps; strictly increasing within each segment |
right_arm_joints / right_hand_joints |
(T,6) / (T,16), source radians |
right_wrist_pose |
(T,6), robot-base JAKA Link_6 flange: meters + rotation-vector radians |
right_tactile_data_fingers |
(T,8,15,3), original uncalibrated sensor readings |
right_tactile_* |
OXT functional areas and sensor/type labels; inspect metadata for exact array names |
sub_task_instruction |
Coarse English instructions inferred from directories |
source_record_sha256, source_episode_id, source_frame_index |
Source lineage identifiers |
source_group |
Original source/version group for every frame |
source_split |
train / test for PlayData; unspecified for other tasks |
raw/ retains source depth, poses, available commands, velocities, efforts and command-presence masks for retained frames. Missing optional hand commands are represented by NaN plus a false presence mask only in raw; canonical numeric state arrays are finite. No control actions were fabricated. Detailed conventions and all exact array names are in the quality report and official structure report.
The eight pads alternate tip/pulp on thumb, index, middle and ring fingers. Fingertip model: PX6AX-GEN1-DP-S2716; finger-pulp model: PX6AX-GEN1-IP-S2516. See official product links and images. Tactile values have no asserted Newton calibration. The flange pose is not a calibrated camera or palm pose. Physical slot placement, axis signs and camera conventions still need collection-owner confirmation.
Curation and check coverage
- Inputs: four source archives, 285 original trajectories and 50,387 records; source archive hashes and all original record CRCs checked.
- Removed 3,068 exact full-record duplicates, including 25 exact play train/test overlaps. This does not prove absence of similar trajectories or other leakage. No invalid observations were discarded; 26 records with missing optional hand commands were recovered.
- Retained 283 source trajectories as 328 continuous segments after real time gaps and internal deduplication breaks. These segments are not 328 independent human demonstrations.
- The initial conversion read back all 94,638 RGB images and spot checked 1,966 depth frames. This selected release retains 94,248 of those RGB images; all retained archive bytes are unchanged from the verified merged release.
- Canonical state finiteness, unique source hashes, source lineage, segment timing and pose conversions passed. Every packaged member also passed ZIP CRC validation, and each packaged Zarr store opened independently.
- Five conversion tests and four upstream tests passed. Numerical checks do not establish physical calibration, exact cross-modal synchronization or task success.
metadata.task=4 now counts the merged task stores. The previous 12 groups were source/version subdivisions. Instructions cover four coarse families: assembly, flipping, opening a box and free interaction. No per-episode success labels were supplied.
Scope requiring review: 25,846 frames retain a source nouse label; task and play contribute 12,464 and 8,814 frames respectively. The four retained tasks are included with provenance labels; the one-trajectory reorient task is excluded at the owner’s request. Source-grounded joint/pad mappings and coarse language labels are documented. Acquisition-level calibration, the training meaning of nouse, richer success labels and future release/license choices remain disclosed limitations; they are not silently inferred. No license is inferred from public availability. PKU and PaXini image provenance is recorded; those marks are not assigned a new license by this dataset card.
Provenance and evidence
- Current selection audit, historical five-task merge audit, original conversion audit (12 source groups), deduplication and segmentation, kinematics validation, source archive hashes, source inventory.
- OXT review materials: metadata, PKU logo, sensor images, hand/tactile maps, 10-second preview video, 20 current VLM pages, 11 diagnostic sheets, and 60 historical montage pages from the previous 12-source-group layout.
- Repository versions and original processing script hashes. The archived local processing audit uses separately retained source files; this Hub snapshot provides the processed outputs and portable read-only QA.
- OXT-QC: https://github.com/nolann030628/OXT-QC/tree/cadbf5e7d1cfcec861a6d58779fc6d6e057f3126
- FTP-1 conventions: https://github.com/michaelyuancb/ftp1-policy/tree/89fa681d6c014cce28300946b7526db808e0b1c1
- Hardware and kinematics reference: https://github.com/tianhaowuhz/3dtacdex/tree/a4a367514ee25c4ad43eb2400cac0141bcb85331
- Reference project: https://3dtacdex.github.io/
Local paths in copied reports are normalized to this repository. Historical checker results are preserved; current contributor metadata and upload state live in OXT_Submission_3dtacdex/metadata.json and submission_status.json.
Merging versions and preserving splits
Assembly combines three assembly_random versions and task__assembly_fr_v1; flip combines four versions; play_data concatenates train then test in one store. Play train occupies frames [0, 7706) and test [7706, 8814). Select the source_split label when building training or evaluation data; do not treat the merged play store as an all-training split. Every source segment is preserved, including at task-version boundaries.
The previous layout remains in repository history at commit a4d582a. Original data values are unchanged.
Excluded task
The owner requested removal of the only task with one trajectory: reorient (195 frames, one segment). The current release has four task archives and 47,124 frames / 328 segments. The other four archives are byte-identical to the verified five-task release. The public preview video now shows a retained assembly trajectory. Earlier conversion records, audit/local_vlm/, audit/historical_vlm_dry_run/ and old source-group montage directories are historical evidence, including the excluded task; they do not add it to the current dataset or count as VLM coverage of the retained tasks. The previous five-task release remains in repository history.
- Downloads last month
- 282