TacGooseBumps (TacGB)

Human demonstration data with visual, tactile, and proprioceptive observations

🌐 Project · 📄 Paper · 🎬 Video · 📦 Data · 📖 Data format

TacGooseBumps (TacGB) is a passive domed film that adds shear-sensitive patterns to the pressure maps of existing normal-only tactile sensors. Tangential contact deforms the domes and redistributes pressure, giving a learning policy additional information about contact without changing the sensor electronics or reconstructing shear forces.

This dataset accompanies our paper. It contains demonstrations collected with bare tactile sensors and TacGB-equipped sensors, supporting research on tactile representations, imitation learning, and vision–touch policies.

810 demonstrations · 4 tasks · 9 task/condition subsets · 263,917 frames · 256 × 256 RGB

Tasks

Task	Collection setup	Conditions	Demonstrations
🔌 USB insertion	Robot teleoperation	Bare, off-the-shelf TacGB, lab-made TacGB	300
💡 Twist-lock lightbulb insertion	Robot teleoperation	Bare, TacGB	110
🥚 Raw egg placement	Handheld iPhUMI	Bare, TacGB	200
✍️ Whiteboard drawing	Handheld iPhUMI	Bare, TacGB	200

The robot recordings include wrist, front, and oblique RGB views with a single 32 × 32 tactile matrix. The handheld recordings include wrist and ultrawide RGB views with two 12 × 32 tactile matrices, one per finger. Both include tool poses, gripper state, timestamps, and task descriptions at approximately 20 Hz.

See the project page for the TacGB hardware design and task demonstrations.

Download

Each subset is packaged as a separate Zarr v2 ZIP archive, so you can download only the tasks and sensor conditions you need. The complete dataset is approximately 91.5 GB.

Archive	Demonstrations	Frames	Size (GB)
usb_bare	100	58,342	21.92
usb_tacgb_off_the_shelf	100	58,744	23.17
USB · lab-made TacGB	100	58,937	22.52
bulb_bare	55	20,443	7.74
bulb_tacgb	55	19,148	7.29
egg_bare	100	13,319	2.36
egg_tacgb	100	18,605	3.40
drawing_bare	100	7,163	1.34
drawing_tacgb	100	9,216	1.72
Total	810	263,917	91.47

The lab-made TacGB USB subset is stored in usb_tacgb_molded.zarr.zip.

Install the reader dependencies:

pip install huggingface_hub "zarr==2.18.3" "numcodecs==0.13.1"

Download one subset:

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="jeffwli/TacGB",
    repo_type="dataset",
    filename="TacGB/egg_bare.zarr.zip",
)

For all subsets, use snapshot_download("jeffwli/TacGB", repo_type="dataset", allow_patterns=["TacGB/*", "SCHEMA.md", "manifest.json", "SHA256SUMS"]). Archive checksums are provided in SHA256SUMS.

Load a demonstration

Read the archive directly without extracting its individual chunk files:

import zarr

with zarr.ZipStore(archive, mode="r") as store:
    root = zarr.open_group(store=store, path="egg_bare.zarr", mode="r")
    episode_ends = root["meta/episode_ends"][:]

    episode = 0
    start = 0 if episode == 0 else int(episode_ends[episode - 1])
    stop = int(episode_ends[episode])
    frames = slice(start, stop)

    rgb = root["data/right_wrist_camera_rgb"][frames]       # (T, 256, 256, 3), uint8
    tactile = root["data/right_tactile_data_gripper"][frames]  # (T, 2, 12, 32)
    pose = root["data/right_wrist_pose"][frames]            # (T, 6): XYZ + rotation vector
    gripper = root["data/right_hand_joints"][frames]        # (T, 1)
    timestamps = root["data/timestamps"][frames]            # (T,), seconds
    episode_id = root["meta/episode_ids"][episode]

print(episode_id, rgb.shape, tactile.shape)

Use the corresponding <archive_name>.zarr group when opening another subset. right_ identifies the single active hand/gripper stream; the handheld tactile array contains both left and right pads.

File organization

TacGB/
  usb_bare.zarr.zip
  usb_tacgb_off_the_shelf.zarr.zip
  usb_tacgb_molded.zarr.zip
  bulb_bare.zarr.zip
  bulb_tacgb.zarr.zip
  egg_bare.zarr.zip
  egg_tacgb.zarr.zip
  drawing_bare.zarr.zip
  drawing_tacgb.zarr.zip
previews/          # Four offline Viser recordings of handheld demonstrations
manifest.json      # Subset and episode index
SHA256SUMS         # SHA-256 checksums for the nine archives
SCHEMA.md          # Fields, units, timing, and coordinate conventions

Each archive contains data/ arrays concatenated across demonstrations and meta/ arrays identifying episode boundaries. The schema describes all fields, including tactile sample stacks and additional recorded signals. These archives are read with Zarr; they do not require the Hugging Face tabular viewer.

Usage notes

• Background faces are pixelated in the released RGB streams. Pixels outside the face masks and all tactile, pose, action, gripper, and timestamp data are unchanged.
• RGB images are provided at 256 × 256. These are processed demonstration streams; preprocessing, timing, and valid tactile sample counts are documented in the schema.
• Tactile matrices encode sensor responses, not calibrated forces in newtons. Gripper state is dimensionless openness for USB/lightbulb and meter-scaled opening for handheld tasks.
• Tool poses use the collection setup's documented reference frame. Apply the appropriate tool-frame mapping when transferring trajectories to different hardware.
• Task descriptions label the overall task. Demonstrations can cover only part of a task; frame-level phase and success labels are not included.
• No predefined train/test split is supplied. Split by demonstration rather than by individual frames when evaluating learning methods.

License

The dataset and accompanying documentation are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt the material, including for commercial purposes, with appropriate attribution, a link to the license, and an indication of any changes. See LICENSE for the full terms.

Citation

If you use the dataset in your research, please cite:

@misc{li2026tacgoosebumps,
  title={TacGooseBumps (TacGB): Retrofitting Normal-Only Tactile Sensors with Shear Encoding for Learning Contact-Rich Manipulation},
  author={Wenjie Li and Binyu Yang and Yuxin Chen and Ambrose Wang and Masayoshi Tomizuka},
  year={2026},
  eprint={2609.34006},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2609.34006}
}

Training and deployment code will be released upon paper acceptance. For questions about the data, please open a discussion.