Humanoid Data is Better
when Universally Sharable

Through TWIST & TWIST2, we have built the most effective & holistic data collection system for humanoid robots.
Now we are towards scaling up humanoid whole-body loco-manipulation data.

TWIST Data Principles

📈 No dataset is too small; small data adds up.

Even a simple, single-task dataset with 50 demonstrations is valuable. As long as you're happy to share it, we are happy to list it here. Every bit of data contributes to the community's collective knowledge.

🔧 Standardized (and affordable) hardware makes faster progress.

We have standardized the hardware (Unitree G1 + Unitree Dex31 Hand + TWIST2 Neck) for unified and scalable data collection. We believe in single robot, maximum performance, instead of cross-embodiment transfer with moderate performance.

🤖 Unified interface enables scalable humanoid policy learning.

We have standardized the state/action format for humanoid policy learning. The state/action are represented as {base xy vel, base z, base roll pitch, base yaw vel, joint positions of body, hand, neck}.

Data Visualizer

Browse and filter TWIST data interactively. Hover over videos and click the text to access Hugging Face links.

Datasets

Project Name Contact Task Data Links Obs Actions # Demos License Notes Citation

📋 Listing Your Data

If you would like your data to be listed here, please provide the following information:

  • Project: The name of the project and link to the project website for additional context.
  • License: The licensing terms under which the data is distributed (e.g., MIT, GPL, CC BY 4.0).
  • Citation: Information on how to properly cite the dataset if used.
  • Contact person: The name and email of the individual responsible for the dataset.
  • Tasks:
    • Metadata: This includes the task name, the observations (e.g., Image, Proprio, Audio, Bimanual), and the actions (e.g., 6 DoF End Effector, Parallel Gripper, Bimanual).
    • Link to Data: A direct link to where each of the data formats can be accessed. For instance, the link for MP4 might be a Zarr link, while the link for a Zarr file might be hosted on the project's webserver. If you're unsure of which data format to use, consider using/extending the TWIST data format.
    • Number of demonstrations: The number of demonstrations in the dataset.

Please send this information to yanjieze@stanford.edu for review and inclusion in our listing.

Disclaimer: We only provide a list of references to existing data for researchers to easily search and find datasets. We do not own or host the datasets listed here. For questions about specific datasets, please reach out to the project owners directly.

📚 TWIST Data Format

We save data (image, state, action). The img is directly saved as .png files. The state and action are saved as .json files.

The file structure of a task is as follows:


TASK_NAME/
    episode_0000/
        rgb/
           000000.png
           000001.png
           ...
        data.json
    episode_0001/
        rgb/
           000000.png
           000001.png
           ...
        data.json
    ...

The data.json file has content like this:

{
  "info": {
    "version": "1.0.0",
    "date": "2025-10-24",
    "author": "YanjieZe",
    "image": {
      "width": 1280,
      "height": 360,
      "fps": 30
    }
  },
  "text": {
    "goal": "humanoid do something",
    "desc": "filtered and segmented teleop data",
    "steps": "filtered based on visibility criteria and segmented into valid sequences"
  },
  "data": [
    {
      "idx": 0,
      "rgb": "rgb/000000.jpg",
      "state_body": [0.0, 0.0, 0.7, ...],
      "state_hand_left": [0.0, 0.1, 0.174, ...],
      "state_hand_right": [0.0, 0.0, 0.0, ...],
      "state_neck": [0.0, 0.0],
      "action_body": [0.0008, -0.019, 0.642, ...],
      "action_hand_left": [0.0, 0.1, 0.174, ...],
      "action_hand_right": [0.0, 0.0, 0.0, ...],
      "action_neck": [-0.218, -0.231]
    }
  ]
}

Each frame contains:

Contact

For questions or to have your dataset listed on this website, please reach out to Yanjie Ze.

Website modified from UMI Data. Thanks to the UMI team for the inspiration.

Data visualizer modified from DROID.