OpenMOSS-Team/robocasa365-lerobot-v3.0

RoboCasa365 LeRobot v3.0

This dataset packages the RoboCasa365 pretraining demonstrations used by EasyWAM in LeRobot v3.0 format. It includes the atomic and composite pretraining corpora with synchronized multi-view RGB video, robot state, actions, rewards, termination flags, and task annotations.

Dataset Summary

Subset	Episodes	Frames	Tasks
Pretrain Atomic	7,356	1,495,313	617
Pretrain Composite	24,687	27,610,913	4,428
Total	32,043	29,106,226	5,045

Both subsets use the PandaOmron embodiment and are recorded at 20 FPS. This release contains the broad pretraining corpus; RoboCasa365 target-task demonstrations are not part of these directories.

Structure

robocasa365-lerobot-v3.0/
├── pretrain-atomic/
├── pretrain-composite/
└── dataset_stats.json

Each subset follows the LeRobot v3.0 layout:

<subset>/
├── data/                  # frame-level Parquet files
├── meta/
│   ├── episodes/          # episode metadata
│   ├── info.json          # schema and dataset totals
│   ├── stats.json
│   └── tasks.parquet      # task descriptions and names
└── videos/                # camera MP4 files

Features

Feature	Type / shape	Description
observation.images.robot0_agentview_left	RGB video, 256×256	Left agent view
observation.images.robot0_agentview_right	RGB video, 256×256	Right agent view
observation.images.robot0_eye_in_hand	RGB video, 256×256	Wrist camera
observation.state	float64[16]	Base pose, relative end-effector pose, and gripper state
action	float64[12]	Base motion, control mode, end-effector delta, and gripper command
next.reward	float32	Next-step reward
next.done	bool	Episode termination flag
annotation.human.task_description	int64	Task-description annotation index
annotation.human.task_name	int64	Task-name annotation index
task_index	int64	Index into meta/tasks.parquet

Videos use H.264 with YUV 4:2:0 pixel format and contain no audio.

Download

hf download OpenMOSS-Team/robocasa365-lerobot-v3.0 \
  --repo-type dataset \
  --local-dir data/robocasa365-lerobot-v3.0

Use with EasyWAM

The EasyWAM data pipeline loads both pretraining subsets, concatenates the left, right, and wrist cameras, and uses a 33-step action/state horizon with 9 decoded video timestamps. The default recipe computes normalization statistics from the combined subsets when training starts.

python scripts/precompute_text_embeds.py task=robocasa_easywam_mot_wan22

NPROC_PER_NODE=8 bash scripts/train_zero1.sh \
  task=robocasa_easywam_mot_wan22

See the EasyWAM RoboCasa365 data guide for configuration details.

Project

• Project page: http://openmoss.ai/EasyWAM/
• Code: https://github.com/OpenMOSS/EasyWAM
• Upstream benchmark: https://github.com/robocasa/robocasa

License and Citation

RoboCasa assets and datasets are distributed under CC BY 4.0. Please retain the original attribution and cite RoboCasa365:

@inproceedings{robocasa365,
  title     = {RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots},
  author    = {Soroush Nasiriany and Sepehr Nasiriany and Abhiram Maddukuri and Yuke Zhu},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}