zhenyuxie-zhzh/EgoTouch_hdf5

EgoTouch HDF5

This repository contains the converted HDF5 release of EgoTouch from the TouchAnything project: a large-scale multi-view tactile dataset for egocentric hand-object interaction.

EgoTouch contains 208 diverse manipulation tasks across 1,891 episodes in indoor and outdoor environments. It provides synchronized egocentric and dual wrist camera videos, bimanual hand pose annotations, and dense continuous pressure maps from wearable tactile sensors.

This HDF5 repository intentionally excludes the local depth/ and pose3d_dynhamr/ directories.

Dataset Contents

Top-level directories:

• Home/, Office/, Outdoor/, Retail/, Workbench/: converted HDF5 episodes grouped by environment and task.
• pose3d/: precomputed 3D hand-pose related files.
• vipe_chest_depth_camera/: auxiliary processed files.
• split.json: official train/validation/test split.

The split file contains:

• train: 1665 episodes
• val: 208 episodes
• test_seen: 208 episodes
• test_unseen: 147 episodes

HDF5 Format

Each episode is stored as an HDF5 file with the following structure:

images/
  chest_color        (T, 480, 640, 3) egocentric RGB
  left_color         (T, 480, 640, 3) left wrist RGB
  right_color        (T, 480, 640, 3) right wrist RGB
hands/
  wilor_left_joint_xyz    (T, 21, 3) left hand pose from WiLoR
  wilor_right_joint_xyz   (T, 21, 3) right hand pose from WiLoR
  wilor_left_valid        (T,) left-hand pose validity mask
  wilor_right_valid       (T,) right-hand pose validity mask
pressure/
  left_pressure_grid      (T, 21, 21) normalized [0, 1]
  right_pressure_grid     (T, 21, 21) normalized [0, 1]
poses/
  chest_pose         (T, 7) camera pose [xyz, quat]
  left_pose          (T, 7) left wrist camera pose
  right_pose         (T, 7) right wrist camera pose
masks/
  glove_masks        (T, N, 480, 640) glove/object masks, N may be 0
  glove_obj_ids      (T, N) object ids for mask channels
  glove_valid_frames (T,) mask-valid frame flags
metadata/
  attrs              task name, trajectory id, fps, frame count, etc.
timestamps           (T,) frame timestamps

T is the number of frames in an episode. Released HDF5 files use 30 FPS and may contain variable-length trajectories.

Loading Example

import h5py
from huggingface_hub import hf_hub_download

repo_id = "zhenyuxie-zhzh/EgoTouch_hdf5"
path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename="Home/play_camel_up/20260410_143747_047.hdf5",
)

with h5py.File(path, "r") as f:
    ego_rgb = f["images/chest_color"]
    left_pressure = f["pressure/left_pressure_grid"]
    print(ego_rgb.shape, left_pressure.shape)

Citation

If you find this dataset or project useful, please cite:

@misc{zhou2026touchanythingdatasetframeworkbimanual,
      title={TouchAnything: A Dataset and Framework for Bimanual Tactile Estimation from Egocentric Video},
      author={Jianyi Zhou and Ziteng Gao and Feiyang Hong and Zirui Liu and Guannan Zhang and Weisheng Dai and Ruichen Zhen and Chuqiao Lyu and Haotian Wu and Yinian Mao and Xushi Wang and Yuxiang Jiang and Wenbo Ding and Shuo Yang},
      year={2026},
      eprint={2605.13083},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.13083},
}

License

This release follows the MIT license used by the TouchAnything project. See the project repository for full license text.