EgoTactile-OXT

This repository contains the Open-X-Tactile conversion of the EgoTactile dataset. It provides aligned egocentric RGB frames and right-hand pressure readings in the OXT Zarr schema.

Contents

• 63 object-level tasks
• 767 trajectories
• 314,100 aligned frames
• EgoTactile/*.zip: one ZIP archive per object-level Zarr task
• supplementary_tactile_note.json: the 162-channel tactile mapping
• qc3/vlm_review.json and qc3/vlm_review_result.json: genuine Ark VLM review results
• qc3/QC3_REVIEW_CN.md: Chinese QC3 findings, evidence and data-provider adjudication
• qc3_montages.zip: all 315 original review images, separate from the small submission package

Each archive expands to one <object>.zarr directory. The data follow the Open-X-Tactile format.

Source and conversion notes

The source dataset contains RGB video, timestamps, and a 256-value right-glove record. The OXT conversion exposes the official 162-location EgoTactile glove layout while preserving the original 256-value vector as an additional field. Pressure values remain in newtons; no training-time clipping or normalization is applied.

The 162-cell OXT layout was independently cross-checked against the auxiliary 256_to_image.py right-hand right_mask_indexed template: every source sensor index and (y, x) cell matches. This is a tactile-pressure topology (137 unique physical indices, with five bend sensors expanded to six cells each), not hand kinematics; it must not be interpreted as right_hand_joints or used to invent joint labels absent from the source release.

The release contains 692 directly paired gloved_hand trajectories and 75 weakly paired bare_hand trajectories. In the latter, the visible task hand and the off-camera instrumented hand perform corresponding actions; visual contact and measured pressure are not guaranteed to match exactly. Use acquisition_mode and pairing_quality to distinguish these cases. Do not interpret successful schema checks as proof of precise physical synchronization for weakly paired sequences.

Each object-level Zarr concatenates multiple source trajectories and participants. meta/episode_ends records their boundaries. The official QC3 montage samples across an entire object, so a change of hand, clothing, or viewpoint between tiles may be a trajectory boundary, not corruption.

Quality control

The conversion's schema precheck passed for all 63 objects. Exact source comparison covered all 314,100 tactile frames and timestamps.

Formal QC3 was run with doubao-seed-2-1-pro-260628 on Ark using the unchanged official OXT checker and default sampling. All 315 image calls returned valid scored responses, with no API-error pages. Object-level judgments were 34 pass, 26 warning, 3 fail; page-level judgments were 277 pass, 35 warning, and 3 fail. The original scores and responses are preserved.

The three model-failed objects are SamyangSpicyChickenCupNoodles (page 1), SprayBottle (page 5), and TennisBall (page 2). On 2026-09-04, the data provider reviewed them and attributed the findings to glove noise/baseline or residual pressure, cross-trajectory montage semantics, and VLM false positives. No data reprocessing is required. The original model scores and fail labels remain unchanged for auditability; see the Chinese review notes.

QC2/4D visualization and the hand-joint-index image are requested exemptions because the public source contains no poses, calibration, or hand joints. Those exemptions are pending organizer approval.

The public release contains 767 complete video/JSON pairs. The paper reports 768 clips; the expected gloved_hand/p001/Sponge/repeat0000 pair is absent from the public release and is not fabricated here.

License and privacy

This conversion is released under the same Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license as the original EgoTactile dataset. Use is limited to non-commercial purposes with appropriate attribution. The original privacy constraints also remain in effect; participant attributes must not be used for identity inference or unintended profiling.

References

• Project: https://egotactile.github.io/
• Source code: https://github.com/Russell-Zeng/EgoTactile
• Source dataset: https://huggingface.co/datasets/HustleHard/EgoTactile