sharinka0715/X-WAM-RoboCasa

X-WAM

Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising

---

Dataset Summary

This is the RoboCasa fine-tuning dataset used to train the X-WAM unified 4D World Action Model. It packages single-arm kitchen manipulation demonstrations into a unified multi-view RGB-D video + low-dimensional state/action format, where each episode provides synchronized RGB videos, depth videos, end-effector proprioception, actions, and language instructions.

• Source benchmark: RoboCasa
• Embodiment: Single-arm manipulator
• Tasks: 24 atomic kitchen manipulation tasks (pick-and-place, open/close doors & drawers, press buttons, etc.)
• Modalities: 3 camera views × (RGB + Depth) + EE proprioception + EE actions + language

Episodes	1,235
Total frames	341,017
Avg. frames / episode	~276
Camera views	3 (2 static + 1 wrist)
Video resolution	256 × 256, H.264
Frame rate	20 fps
Instructions / episode	1
Total size	~5.1 GB

---

Dataset Structure

.
├── data/                          # Per-episode low-dim states, actions & metadata (JSON)
│   └── chunk-{0000..0001}/
│       └── episode_{id}.json
├── video/                         # RGB videos (.mp4, H.264)
│   ├── robot0_agentview_left/chunk-{id}/episode_{id}.mp4
│   ├── robot0_agentview_right/chunk-{id}/episode_{id}.mp4
│   └── robot0_eye_in_hand/chunk-{id}/episode_{id}.mp4
├── depth/                         # Depth videos (.mp4, H.264; same layout as video/)
│   ├── robot0_agentview_left/...
│   ├── robot0_agentview_right/...
│   └── robot0_eye_in_hand/...
└── metadata.json                  # { "chunk-xxxx/episode_xxxxxxx": num_frames }

metadata.json maps each episode key to its number of frames, e.g.:

{
  "chunk-0000/episode_0000000": 413,
  "chunk-0000/episode_0000001": 412
}

Camera Views

View	Type	Description
robot0_agentview_left	static	Left third-person view
robot0_agentview_right	static	Right third-person view
robot0_eye_in_hand	dynamic	Wrist-mounted (eye-in-hand) view

Episode Schema (data/.../episode_{id}.json)

Field	Type	Description
num_frames	int	Number of frames N in the episode
instructions	list[str]	Natural-language task description (1 per episode)
observations	dict	Per-camera video references (see below)
proprios	dict	Per-frame proprioceptive state, length N
actions	dict	Per-frame actions, length N-1

Each entry in observations[<camera>] points to the corresponding RGB/Depth clip:

{
  "type": "static",
  "rgb_path":   "video/robot0_agentview_left/chunk-0000/episode_0000000.mp4",
  "depth_path": "depth/robot0_agentview_left/chunk-0000/episode_0000000.mp4",
  "start": 0,
  "end": 413,
  "fps": 20.0
}

State & Action Spaces (single-arm)

Group	Key	Dim	Description
proprios	left_ee_pos	3	EE position (xyz), absolute
proprios	left_ee_rotm	9	EE rotation matrix (flattened 3×3), absolute
proprios	left_gripper_pos	1	Gripper opening
actions	left_ee_pos	3	Target EE position
actions	left_ee_rotm	9	Target EE rotation matrix
actions	left_gripper_pos	1	Target gripper opening
actions	raw_actions	7	Raw environment action

---

Usage

# Please refer to the code repository for full data loading, training and evaluation scripts:
# https://github.com/sharinka0715/X-WAM

---

Source & Attribution

This dataset is derived from the RoboCasa benchmark, re-rendered with multi-view RGB-D and re-packaged into the unified X-WAM format. Please also cite and comply with the license of the original RoboCasa benchmark when using this data.

---

Citation

If you use this dataset, please cite X-WAM:

@article{guo2026xwam,
  title={Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising},
  author={Guo, Jun and Li, Qiwei and Li, Peiyan and Chen, Zilong and Sun, Nan and Su, Yifei and Wang, Heyun and Zhang, Yuan and Li, Xinghang and Liu, Huaping},
  journal={arXiv preprint arXiv:2604.26694},
  year={2026}
}

License

This dataset is released under the Apache License 2.0, subject to the terms of the upstream RoboCasa benchmark.