🌍 JEPA-WMs Datasets

Robotics trajectories for world model training 🤖

Meta AI Research, FAIR

This 🤗 HuggingFace repository hosts datasets for training JEPA-WM world models.
👉 See the main repository for training code and pretrained models.

👁️ Preview Images: To view example images in the Dataset Viewer above, select a dataset configuration (e.g., metaworld, pusht) and click "Run query".

---

📦 Downloading Data

Use the download script from the main repository:

# Download all datasets
python src/scripts/download_data.py

# Download specific dataset(s)
python src/scripts/download_data.py --dataset pusht pointmaze wall

# List available datasets
python src/scripts/download_data.py --list

---

📋 Available Datasets

Dataset	Description	Format
🏭 metaworld	Tabletop manipulation (42 tasks)	.mp4 + .parquet
🏠 robocasa	Kitchen manipulation	.hdf5
🦾 franka_custom	Real Franka robot (3 views)	.h5 per episode
🔵 pusht	Push-T block pushing	.zip 📦
🚪 wall	Point navigation through doors	.zip 📦
🧩 point_maze	Point navigation in mazes	.zip 📦

💡 The pusht, wall, and point_maze datasets are sourced from DINO-WM and re-hosted here for convenience.

---

📚 Dataset Details

🏭 Metaworld

Tabletop robotic manipulation across 42 different tasks.

Field	Shape	Description
observation	224×224 RGB	Rendered observation image
state	39-dim	Full state vector
action	4-dim	End-effector action
reward	scalar	Task reward
task	string	Task name (e.g., "drawer-open")

🏠 RoboCasa

Kitchen manipulation with multiple camera views.

Field	Shape	Description
eye_in_hand	256×256 RGB	Eye-in-hand camera
leftview	256×256 RGB	Left view camera
action	12-dim	Robot action
state_*	various	State observations

🦾 Franka Custom

Real Franka robot with 3 camera views.

Field	Shape	Description
exterior_image_1_left	480×640 RGB	Exterior camera 1
exterior_image_2_left	480×640 RGB	Exterior camera 2
wrist_image_left	480×640 RGB	Wrist-mounted camera
cartesian_position	6-dim	End-effector pose
joint_position	7-dim	Joint angles
gripper_position	scalar	Gripper state

🔵 Push-T

Block pushing task from the Push-T benchmark.

Field	Shape	Description
observation	224×224 RGB	Rendered observation
state	5-dim	Block + agent state
action	2-dim	Relative position action
velocity	2-dim	Agent velocity

🚪 Wall

Point navigation through walls with doors.

Field	Shape	Description
observation	224×224 RGB	Rendered observation
state	2-dim	Position (x, y)
action	2-dim	Movement action
door_location	scalar	Door y-position
wall_location	scalar	Wall x-position

🧩 Point Maze

Point navigation in procedural mazes.

Field	Shape	Description
observation	224×224 RGB	Rendered observation
state	4-dim	Position + velocity
action	2-dim	Movement action

---

📁 Repository Structure

.
├── 📄 README.md
├── 📄 pyproject.toml
├── 📂 scripts/                    # 🛠️ Utility scripts
│   ├── convert_to_hf.py           #   Convert raw → parquet
│   ├── visualize.py               #   Visualize converted data
│   └── upload_to_hf.py            #   Upload to HuggingFace
├── 📂 metaworld/
│   ├── hf_data/                   # Example parquet (for dataset viewer)
│   └── data/                      # Raw parquet files
├── 📂 robocasa/
│   ├── hf_data/                   # Example parquet (for dataset viewer)
│   └── combine_all_im256.hdf5     # Raw HDF5
├── 📂 franka_custom/
│   ├── hf_data/                   # Example parquet (for dataset viewer)
│   └── data/                      # Raw H5 files (per episode)
├── 📂 pusht/
│   ├── hf_data/                   # Example parquet (for dataset viewer)
│   └── pusht_noise.zip            # Raw data (zipped)
├── 📂 wall/
│   ├── hf_data/                   # Example parquet (for dataset viewer)
│   └── wall_single.zip            # Raw data (zipped)
└── 📂 point_maze/
    ├── hf_data/                   # Example parquet (for dataset viewer)
    └── point_maze.zip             # Raw data (zipped)

---

🛠️ Development Scripts

These scripts are for dataset maintainers and developers.

🔄 Convert Raw Data to Parquet

# Analyze dataset structure
python scripts/convert_to_hf.py --dataset metaworld --analyze

# Convert episode 0 (default)
python scripts/convert_to_hf.py --dataset metaworld --convert
python scripts/convert_to_hf.py --dataset pusht --convert
python scripts/convert_to_hf.py --dataset wall --convert
python scripts/convert_to_hf.py --dataset point_maze --convert
python scripts/convert_to_hf.py --dataset robocasa --convert
python scripts/convert_to_hf.py --dataset franka_custom --convert

# Convert specific episode with options
python scripts/convert_to_hf.py --dataset wall --convert --episode 5 --max-frames 50

👀 Visualize Converted Data

# Display frames in matplotlib window
python scripts/visualize.py --dataset metaworld
python scripts/visualize.py --dataset pusht

# Save visualization to file
python scripts/visualize.py --dataset point_maze --num-frames 12 --save output.png

# Print dataset info only
python scripts/visualize.py --dataset robocasa --info-only

☁️ Upload to HuggingFace

# Upload a single file
python scripts/upload_to_hf.py --file robocasa/hf_data/train-00000-of-00001.parquet

# Upload an entire folder
python scripts/upload_to_hf.py --folder franka_custom --message "Add franka_custom data"

# Upload all parquet files
python scripts/upload_to_hf.py --all

---

📄 License

This dataset is released under the CC-BY-4.0 License.

---

📚 Citation

If you find these datasets useful, please consider giving a ⭐ and citing:

@misc{terver2025drivessuccessphysicalplanning,
      title={What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?},
      author={Basile Terver and Tsung-Yen Yang and Jean Ponce and Adrien Bardes and Yann LeCun},
      year={2025},
      eprint={2512.24497},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.24497},
}

---

Made with ❤️ by Meta AI Research, FAIR