---
license: cc-by-4.0
language:
  - en
pretty_name: Humanoid Touch Dream — Open-X-Tactile
tags:
  - robotics
  - humanoid
  - tactile
  - egocentric
  - open-x-tactile
---

# Humanoid Touch Dream

Humanoid Touch Dream (HTD) provides **871 robot demonstrations across six tasks**, collected with a **Unitree G1 humanoid** equipped with **Inspire DFTP or F1 dexterous hands**. The dataset combines egocentric RGB images, tactile sensing, hand motor-force feedback, robot joint states, wrist poses, and language instructions for learning contact-rich manipulation.

This release uses the **[Open-X-Tactile](https://open-x-tactile.github.io/) data format**, with one Zarr v2 episode per ZIP archive.

**[Project](https://humanoid-touch-dream.github.io/) · [Paper](https://arxiv.org/abs/2604.13015) · [Browse episodes](https://huggingface.co/datasets/chrisyrniu/HTD/tree/main/HTD)**

## Dataset at a glance

| | |
| --- | --- |
| Episodes | 871: 760 DFTP and 111 F1 |
| Frames | 994,778 |
| Download size | Approximately 388 GB (361 GiB) |
| Images | 640 × 480 RGB |
| Recording rate | Approximately 30 Hz; use the recorded timestamps |
| Format | Zarr v2, packaged as one ZIP per episode |
| Language | English task instructions |
| License | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) |

| Task folder | Hand | Episodes | Task instruction |
| --- | --- | ---: | --- |
| `book` | DFTP | 162 | Pick up the book from the table and place it on the bookshelf. |
| `insert_t` | DFTP | 208 | Pick up the T-shaped block and insert it into the matching slot. |
| `scoop` | DFTP | 163 | Pick up the scoop, scoop litter from the box, and empty it into the trash bin. |
| `tea` | DFTP | 77 | Pick up both cups from the counter, carry them to the nearby table, and set them down. |
| `towel` | DFTP | 150 | Fold the towel on the table. |
| `lab` | F1 | 111 | Clean the lab. |

Original split labels are preserved in the manifest: **767 training, 57 validation, and 47 unspecified episodes**. Choose how to use the unspecified episodes before constructing a training or evaluation split.

## Download and read an episode

Episodes can be downloaded independently. To start with a single tea demonstration, install the reader dependencies:

```bash
pip install huggingface_hub "zarr>=3,<4" numcodecs imagecodecs imagecodecs-numcodecs
```

The example below downloads the manifest and one episode, then reads its first RGB frame directly from the ZIP. It does not extract the archive or load the whole episode into memory.

```python
import json
from pathlib import Path

import zarr
from huggingface_hub import hf_hub_download
from imagecodecs.numcodecs import register_codecs

repo_id = "chrisyrniu/HTD"
manifest_path = hf_hub_download(
    repo_id, "dataset_manifest.json", repo_type="dataset"
)
manifest = json.loads(Path(manifest_path).read_text())
entry = next(
    episode for episode in manifest["episodes"]
    if episode["archive"].startswith("HTD/InspireHandDFTP/tea/")
)
archive_path = hf_hub_download(
    repo_id, entry["archive"], repo_type="dataset"
)

# Register JPEG / JPEG XL codecs before reading image arrays.
register_codecs(force=False, verbose=False)
with zarr.storage.ZipStore(archive_path, mode="r") as store:
    episode = zarr.open_group(
        store=store, path=entry["zarr_directory"], mode="r"
    )
    rgb = episode["data/camera_ego_rgb"][0]  # uint8, (480, 640, 3)
    instruction = episode["data/sub_task_instruction"][0]
    print(rgb.shape, instruction)
```

Change the task prefix to select another task, such as `HTD/InspireHandF1/lab/`. The manifest includes each archive's path, size, SHA-256 checksum, split, and frame count. For reproducible experiments, pass the same repository commit ID as `revision` to both downloads.

To download the complete dataset:

```bash
hf download chrisyrniu/HTD --repo-type dataset --local-dir htd_data
```

For additional download options, see the [Hugging Face download guide](https://huggingface.co/docs/huggingface_hub/guides/download). For training integration, see the [FTP-1 data-processing pipeline](https://github.com/michaelyuancb/ftp1-policy/blob/main/data_processing/README.md). The [OXT-QC reader](https://github.com/nolann030628/OXT-QC) can also open extracted episodes.

## File organization

```text
HTD/
├── InspireHandDFTP/
│   ├── book/*.zip
│   ├── insert_t/*.zip
│   ├── scoop/*.zip
│   ├── tea/*.zip
│   └── towel/*.zip
└── InspireHandF1/
    └── lab/*.zip
dataset_manifest.json
metadata.json
supplementary_note.json
supplementary_joints_note.json
notes/
README.md
LICENSE-DATA.txt
```

Each ZIP contains one uniquely named `.zarr/` directory. Arrays under `data/` share the same time dimension; `meta/` contains episode boundaries, and `provenance/` preserves source information and original numeric streams. Additional quality reports are available in `oxt_check_output/`.

## Available signals

**Images.** Every episode includes a 640 × 480 ego camera stream. Tea episodes contain only this camera; the other tasks also include left and right wrist cameras. The ego stream uses the selected left stereo view, stored as lossless JPEG XL relative to the decoded source pixels. Wrist streams preserve the original JPEG encoding. Image resizing for training is left to the training pipeline.

**Robot state.** Each side has seven arm joints and six hand joints. `supplementary_joints` provides 15 measured leg and waist joint positions in radians; [supplementary_joints_note.json](supplementary_joints_note.json) defines their order. Original commands, including forward, lateral, and yaw velocity commands, are preserved in `provenance/source_streams/actions`. FTP-1 derives action targets from future observed states rather than these archived commands.

**DFTP tactile sensing.** Each hand provides 1,062 piezoresistive readings, stored as sensor counts divided by 4,095. Native patches are grouped into 11 Open-X-Tactile MTTS surface areas. Padding masks identify the added cells used to combine differently sized patches.

**F1 tactile sensing.** Each hand provides 37 capacitive features. Each fingertip contains `[normal_force/20, tangential_force/20, sin(direction), cos(direction), proximity]`; three palm regions contain the four force/direction features each. Proximity is a normalized sensor reading, not a distance in metres. The direction pair `(0,0)` preserves the SDK's no-contact flag and may occur alongside nonzero force readings.

**Hand motor-force feedback.** Both hand types include six motor-force feedback channels per hand, stored separately from surface tactile readings. These retain the vendor's `force_act/1000` scale. Although assigned to MTTS torque areas, they are **not calibrated joint torques in N·m**.

**Language.** `data/sub_task_instruction` contains the task instruction repeated for every frame in an episode. Instructions describe the intended task; they do not provide intermediate subtask boundaries or per-episode success labels.

See [sensor descriptions](notes/sensor_descriptions.json) for exact tactile shapes, channel order, units, area mappings, and padding definitions.

## Using poses and timestamps

Wrist poses and the ego camera pose are expressed relative to the robot's **moving torso**, using XYZ positions in metres followed by rotation vectors in radians. The standard fields `left_wrist_pose`, `right_wrist_pose`, and `camera_ego_pose` have explicit aliases `left_wrist_to_torso_pose`, `right_wrist_to_torso_pose`, and `camera_ego_to_torso_pose_estimated`. These contain identical values.

The camera mount transform and camera intrinsics are **estimated**. The camera pose is constant relative to the torso even while the robot moves. The auxiliary `torso_pose_to_robot_ground_frame` uses a frame estimated from the current feet configuration; it does not provide a persistent room/world trajectory. Use the [pose and camera notes](notes/pose_and_camera.json) when interpreting coordinates or adapting the data to a model that expects world-frame poses.

Timestamps come from the host camera-read times. Modalities share exported rows, but independent sensor acquisition times are unavailable. Duplicate timestamps have been removed consistently across modalities. There are **165 preserved timing gaps across 102 episodes**, including eight gaps of approximately two seconds. Use timestamps when building temporal windows, and avoid treating a window spanning a long gap as a uniformly sampled sequence. [Timing details](notes/timing_warnings.json) identify the affected intervals.

## Citation and contact

If you use this dataset, please cite [Learning Versatile Humanoid Manipulation with Touch Dreaming](https://arxiv.org/abs/2604.13015):

```bibtex
@article{niu2026touchdreaming,
  title={Learning Versatile Humanoid Manipulation with Touch Dreaming},
  author={Niu, Yaru and Fang, Zhenlong and Chen, Binghong and Zhou, Shuai and
          Senthilkumaran, Revanth Krishna and Zhang, Hao and Chen, Bingqing and
          Qiu, Chen and Tseng, H. Eric and Francis, Jonathan and Zhao, Ding},
  journal={arXiv preprint arXiv:2604.13015},
  year={2026}
}
```

The data is released under **[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)**. Credit the dataset authors, link the license, and indicate any changes. See [LICENSE-DATA.txt](LICENSE-DATA.txt) for the full terms.

**Institutions:** Carnegie Mellon University; Bosch Center for Artificial Intelligence  
**Contact:** [yarun@andrew.cmu.edu](mailto:yarun@andrew.cmu.edu)