TouchScale for Open-X-Tactile
The TouchScale 100 h release, converted to the FTP-1 zarr format for the Open-X-Tactile (OXT) dataset collection. The data format (keys, coordinate frames, tactile function areas) is documented on the OXT website.
| Tasks | 920 (one zarr per task) |
| Trajectories | 15,488 |
| Frames | 3,521,366 at 10 Hz (97.8 h) |
| Type / label | human, ego |
| Tactile sensor | TachinTechEIG012: bimanual full-hand gloves, 880 taxels per hand, normal + shear |
Files
TouchScale_OXT/
<task_name>_<hash>.zarr.zip # one FTP-1 zarr per task, zipped without recompression
README.md
Each zip holds one <task>.zarr/ folder: unzip it, or open it in place with zarr (2.x or 3.x):
import zarr
store = zarr.storage.ZipStore("TouchScale_OXT/<task>.zarr.zip", mode="r")
root = zarr.open_group(store, mode="r", path="<task>.zarr")
ends = root["meta/episode_ends"][:] # episode boundaries
rgb = root["data/camera_ego_rgb"] # (T, 224, 224, 3) uint8
Contents of each zarr (data/)
| Key | Shape | Notes |
|---|---|---|
camera_ego_rgb, left_wrist_camera_rgb, right_wrist_camera_rgb |
(T, 224, 224, 3) uint8 | full 1280×720 frames resized to 224×224 (no crop) |
camera_ego_pose |
(T, 6) | head-camera pose T_world_cam as (x, y, z, rotation vector); OpenCV camera axes; world = first frame of the episode; estimated with ViPE (RGB-D, calibrated intrinsics) |
{left,right}_wrist_pose |
(T, 6) float32 | estimated wrist pose T_world_wrist as (x, y, z, rotation vector), same world frame as camera_ego_pose; short detection gaps interpolated (see below) |
{left,right}_hand_joints |
(T, 21) float32 | 21 MANO joint angles in radians, computed from the estimated hand keypoints with FTP-1's get_hand_joints_mano_single_hand; short detection gaps interpolated |
{left,right}_hand_joints_idx |
(T, 21) int64 | FTP-1 MANO joint ids [1,26,2,3,6,7,8,9,11,12,13,14,16,17,18,19,21,22,23,24,27] |
{left,right}_tactile_data_<group> |
(T, N, rows, cols, 3) float32 | last dim = (normal, shear_x, shear_y), values as in the TouchScale release |
{left,right}_tactile_area_<group> |
(T, N) int64 | OXT tactile function-area id of each pad |
{left,right}_tactile_sensor_<group>, ..._type_<group> |
(T,) | TachinTechEIG012, matrix |
sub_task_instruction |
(T,) | task instruction |
timestamps |
(T,) float64 | Unix time in seconds |
Tactile groups and their function-area ids:
| Group | Pads | Area ids |
|---|---|---|
distal_thumb |
thumb tip | 0 |
distal_fingers |
index, middle, ring, little tips | 1, 2, 3, 4 |
palm |
thenar + hypothenar pads side by side (13 × 16) | 5 |
proximal |
thumb middle, index / middle / ring base | 6, 11, 12, 13 |
middle_fingers |
index, middle, ring, little middle segments | 7, 8, 9, 10 |
The glove has no little-finger base pad, so area 14 is unused. Left-glove pads are stored as the mirror image of the right-glove layout.
Processing notes
- Episodes whose head pose failed a gyroscope consistency check (1 s rotation error median > 2° or p95 > 8°) were dropped: 236 of 15,324 episodes (1.5%).
- Hand state (
wrist_pose,hand_joints) is estimated from the head camera, not measured:- Each 10 Hz head frame goes through WiLoR (MANO). WiLoR's hand detector misses the tactile gloves, so we fine-tuned it on glove frames pseudo-labelled with OWLv2.
- The 3-D hand position is fitted to the 2-D keypoints with the calibrated head intrinsics, then scaled to match the registered head depth at the palm. This depth step was applied on 99.1% (left) and 97.6% (right) of detected frames; the rest keep the 2-D fit.
- Wrist frame: origin at the MANO wrist keypoint; with the hand flat, palm down, Z points along the fingers, X to the right and Y down, for both hands (OXT ego convention). It is mapped to the world with the ViPE head pose.
- Hands were detected in 98.8% (left) and 99.1% (right) of the original frames. The hand keys contain no NaN: gaps of up to 2 s (20 frames) are interpolated (position linearly, rotation with slerp, joint angles along the shorter arc) and gaps at an episode's start or end take the nearest detected frame. Where a hand is missing for longer than 2 s, the episode is cut there and the gap frames are dropped; parts shorter than 2 s are dropped too. 408 of the 15,088 episodes had such a gap and became 808 episodes; 18,780 frames (0.53%) were dropped in all, leaving 15,488 episodes and 3,521,366 frames. 0.84% of the kept hand-frames are interpolated. Each part's first head-camera frame is the world origin.
- FTP-1's adduction angles are a signed arccos, so for curled fingers they can sit near ±π and flip sign between frames. This follows FTP-1's definition and was kept.
- All 920 task zarrs pass the OXT-QC metadata precheck (rerun on the final data).
License and citation
Released under CC BY-NC 4.0, the same license as TouchScale: free for non-commercial research and education with attribution. For commercial use, please contact 2077AI.
If you use this data, please cite TouchScale:
@article{li2026touchscale,
title = {TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning},
author = {Li, Dayou and Wang, Hao and Yang, Qianqian and Zhu, Zihao and Fang, Haoquan and Zeng, Ziyao and Han, Yan and Wang, Zihan and Wang, Yan and Huang, Baoru and Wang, Dilin and Shimada, Kenji and Luo, Yiyue and Li, Manling and Lv, Teresa and Mukadam, Mustafa and Ranjan, Rakesh and Zhang, Ruohan and He, Qi and Liu, Changliu and Chen, Xu and Pavone, Marco and Liu, Bangya and Li, Jiachen and Tomizuka, Masayoshi and Fan, Zhiwen},
journal = {arXiv preprint arXiv:2610.10288},
year = {2026}
}
- Downloads last month
- 300
Paper for 2077AIDataFoundation/TouchScale_OXT
Paper • 2610.10288 • Published • 8