---
license: h-research
pretty_name: WAM-0.1
tags:
- robotics
- panda
- robosuite
- robocasa
- h-jepa
- world-model
---

# WAM-0.1 Dataset

WAM-0.1 is an image-and-action dataset for learning Panda robot dynamics and world representations.

Created by **Atom Ctrl** and **Aditya Panigrahi**.

Website: https://atomctrl.com

GitHub: https://github.com/adityapp1213/Wam-0.1

Hugging Face: https://huggingface.co/datasets/atomctrl/wam-0.1

## Dataset overview

The dataset contains two aligned learning levels.

![Dataset overview](https://cdn-uploads.huggingface.co/production/uploads/68dc1570c7fc381a7bae2aaa/JxfpN9VA5M5vEnCuEKV-6.png)

### Level 0

One action predicts the next observation:

`input image + 7D action -> next image`

### Level 1

An action chunk predicts a future subgoal:

`start image + action chunk -> end image`

Level 1 samples correspond directly to sequences of Level 0 transitions.

## Data sources

- Official RoboSuite Panda tasks
- Randomized Panda trajectories
- Official RoboCasa demonstrations
- RGB and depth images
- Robot state and Panda actions
- Level 0 transitions and Level 1 action chunks

## Visual examples

### Normal trajectories

![Normal trajectory](https://cdn-uploads.huggingface.co/production/uploads/68dc1570c7fc381a7bae2aaa/5NM8_MREYvVHvie2txJo-.png)

### Randomized trajectories

![Random trajectory](https://cdn-uploads.huggingface.co/production/uploads/68dc1570c7fc381a7bae2aaa/VvZIjkTKYBmxTBo_QuqHF.png)

### RoboCasa trajectories

![RoboCasa trajectory](https://cdn-uploads.huggingface.co/production/uploads/68dc1570c7fc381a7bae2aaa/R_Kg1TxijlVv9Ona0pqRt.png)

### Additional examples

![Additional trajectory](https://cdn-uploads.huggingface.co/production/uploads/68dc1570c7fc381a7bae2aaa/FECjQ5dwtG3gs9K2Rh4AW.png)

## Core tasks

- Lift
- Stack
- Door
- Wipe
- PickPlaceBread
- PickPlaceMilk
- PickPlaceCereal
- PickPlaceCan
- PickPlaceSingle
- NutAssemblySquare

RoboCasa contributes additional official kitchen and household manipulation tasks.

## Dataset layout

    data/
    └── final/
        ├── level_0/
        │   ├── normal/
        │   ├── random/
        │   └── robocasa/
        └── level_1/
            ├── normal/
            ├── random/
            └── robocasa/

## Google Colab example

This example downloads only one shard instead of the complete dataset.

    !pip -q install huggingface_hub

    from huggingface_hub import hf_hub_download

    shard = hf_hub_download(
        repo_id="atomctrl/wam-0.1",
        repo_type="dataset",
        filename="shards/lift-normal.tar",
        local_dir="/content/wam-test"
    )

    print(shard)

Extract the shard:

    import tarfile
    from pathlib import Path

    extract_dir = Path("/content/wam-test/extracted")
    extract_dir.mkdir(parents=True, exist_ok=True)

    with tarfile.open(shard, "r") as tar:
        tar.extractall(extract_dir)

    print("Extracted to:", extract_dir)

## Intended use

WAM-0.1 is intended for H-JEPA training, image-based world models, action-conditioned prediction, robot dynamics learning, subgoal prediction, and robotics planning research.

## Download

    pip install -U huggingface_hub

    hf download atomctrl/wam-0.1 \
      --repo-type dataset \
      --local-dir wam-0.1

## Links

- GitHub code: https://github.com/adityapp1213/Wam-0.1
- Dataset: https://huggingface.co/datasets/atomctrl/wam-0.1
- Website: https://atomctrl.com

## Citation

    @dataset{panigrahi_wam_0_1,
      author    = {Aditya Panigrahi},
      title     = {WAM-0.1: World Action Model Dataset},
      year      = {2026},
      publisher = {Atom Ctrl},
      url       = {https://huggingface.co/datasets/atomctrl/wam-0.1}
    }