TactX

TactX — Cross-Sensor Tactile Manipulation Dataset

544 teleoperated Franka episodes: 3 tactile sensors × 4 tasks, with wrist + third-person RGB, robot state, actions, and raw tactile readings.

Paper and code: https://tactx-project.github.io/

Layout

teleop_data/<sensor>/<task>/episode_XXXX.hdf5

sensors: daimon, eflesh, flexitac
tasks:   object_pnp, plug_insertion, spam_reorientation, board_wipe

Train and val share a folder; the original split (482 / 62) is in val_split.json.

HDF5 schema

Raw tactile dimension D, per finger:

Loading

import h5py
with h5py.File("teleop_data/eflesh/object_pnp/episode_0000.hdf5") as f:
    state = f["observation.state"][:, :8]
    tac0  = f["observation.tac0"][:]          # (T, 15)

from huggingface_hub import snapshot_download
snapshot_download("TactX/TactX", repo_type="dataset",
                  allow_patterns="teleop_data/eflesh/**")

Citation

@misc{park2026tactxlearningsharedtactile,
      title={TactX: Learning Shared Tactile Representations Across Diverse Sensors}, 
      author={Junsung Park and Sachin Bhadang and Carmelo Sferrazza and Sha Yi and Xiaolong Wang},
      year={2026},
      eprint={2606.31236},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.31236}, 
}