Uni-VLaT

Uni-VLaT whole body tactile dataset

Uni-VLaT contains 289 demonstrations and 379,397 retained frames from a Unitree G1 without physical hands. This private staging release is being uploaded for Open-X-Tactile. The OXT archives are incomplete until oxt_v1_20261011/UPLOAD_MANIFEST.json is present. Upload completion does not mean OXT acceptance or completion of semantic and coordinate review.

The existing root dataset.zip is the earlier release and is retained unchanged. The new OXT Zarr data are under oxt_v1_20261011/; do not confuse the two formats.

Tasks and archives

Each task folder contains independently extractable ZIPs, one per episode, and its supplementary joint and tactile notes. Each ZIP contains one zarr/*.zarr store and a SHA256SUMS manifest. Extract each ZIP to its own directory and run sha256sum -c SHA256SUMS. The final upload manifest records each ZIP's SHA-256 and size. No train/validation split is introduced. Basket Loading and Composed Cleanup use the retained V2 datasets. Exact instructions are in task_inventory.json.

Both ego RGB eyes are preserved at 640×480, with the cleaned dataset timeline at 50 Hz. data/camera_ego_rgb is the left eye; data/supplementary_ego_right_rgb is the right eye, not a wrist camera. Raw hardware timestamps and uninterrupted physical continuity cannot be inferred from the cleaned timeline. Standard arm joint arrays have seven dimensions per side; supplementary joints contain 12 leg and three waist joints. Physical hand arrays are absent. Original numerical columns are preserved under univlat_source; the 64-dimensional SONIC action token is not a joint-angle vector.

Tactile signals and quality

The release retains 256 wire-order values for each of the vest, left sleeve and right sleeve. No new pooling, deadband, normalization or force conversion is applied. The vest has 112 supplier-defined effective channels; 256 stored values do not establish 256 active vest sensors. Values are counts, not calibrated Newtons or pressure units; acquisition-side baseline correction may have occurred. Supplier mapping and sensor figures are in documentation/. Legacy sensor names ending in UNCONFIRMED are retained to match the arrays; the manufacturer/model mapping is documented separately.

All 289 episodes passed the adapted structural hard checks. At conversion, decoded RGB frames were checked against Zarr readback and numerical mappings were checked for equality. Full semantic/privacy review, external VLM review and calibrated 4D coordinate validation remain incomplete. QC warnings are not erased. Sweep episodes 16–24 have an all-zero left sleeve; Hug episodes 7, 31 and 37 have some nonzero legacy hand teleoperation targets, not evidence of physical hands. These cases are retained and disclosed.

The chest grid display flip and vest right-arm patch ordering discrepancy remain documented in documentation/mapping_comparison.json. Sleeve circumferential placement is uncalibrated. No speculative anatomical reorder has been applied.

Camera reference

Camera visualization uses the default head_camera parameters from the SONIC g1_29dof_old.xml simulation model: position [0.06, 0, 0.45] m relative to torso_link, and MuJoCo Euler angles [0, -0.8, -1.57] rad. camera_reference.json provides the source and the corresponding optical-frame transform. This model-based reference is separate from the recorded arrays. Recorded end-effector poses use fixed-base forward kinematics; world-frame camera trajectories are not included.

License and attribution

Contributor-owned dataset content is licensed under CC BY 4.0. Attribute the Uni-VLaT project and its authors, link the license, and identify modifications. This license does not relicense third-party vendor materials, institutional marks, software or personal likeness rights. Vendor figures are shared with contributor-confirmed public-use permission; no unrestricted relicensing of them is claimed.

Project: https://ggkiller-air.github.io/Uni-VLaT/

Paper: Uni-VLaT: Whole-Body Tactile Adaptation of VLA Policies for Humanoid Loco-Manipulation.

Authors: Zihao Wang, Shutong Liu, Siqi Zheng, Liu Cao, Ruoqu Chen, Rundong Liu, Yanchao Yang, Mengdi Xu.

Contact: Liu Cao, [email protected].