Whalswp/robocasa_pretrain_human300_v3

RoboCasa Pretrain Human 300 — LeRobot v3

This is one LeRobot v3 dataset assembled from the 300 task-folder datasets in Whalswp/robocasa_v3_official_presliced. The direct source is the seed-0, 100-episode-per-task-folder slice of the RoboCasa pretrain_human300 dataset soup. The merge preserves the source order: all 65 atomic task folders first, followed by all 235 composite task folders.

The upstream simulator and dataset are provided by RoboCasa / RoboCasa365.

Dataset summary

Item	Value
Source task folders	300
Atomic / composite folders	65 / 235
Episodes	30,000
Frames	27,815,381
FPS	20
Per-folder used description references	4,615
Globally unique descriptions	4,610
Data Parquet files	300
Video files	954
Video bytes	51,480,763,654
Robot type	PandaOmron
LeRobot codebase version	v3.0

total_tasks=4,610 in meta/info.json means globally unique language descriptions, not the number of RoboCasa task folders. Five language strings occur in two source folders and therefore share one global task_index. The 300 folder-level tasks are tracked separately by source_task_index.

Features

• observation.images.robot0_agentview_left: H.264 video, 256×256 RGB
• observation.images.robot0_agentview_right: H.264 video, 256×256 RGB
• observation.images.robot0_eye_in_hand: H.264 video, 256×256 RGB
• observation.state: float64[16]
• action: float64[12]
• next.reward: float32
• next.done: bool
• annotation.human.task_description: global language-description index
• annotation.human.task_name: source task-folder index, 0..299

All videos are 20 FPS, H.264, yuv420p, without audio.

Source slicing and merge

The direct source was already presliced from the 32,043-episode parent collection by selecting 100 episodes per task folder with seed 0. This merge does not resample episodes.

The dataset was assembled with LeRobot 0.6.1 using:

• concatenate_data=false
• concatenate_videos=false
• atomic folders before composite folders
• no video re-encoding
• one output Parquet shard per source task folder

Data and video shards remain separate. Parquet files were rewritten only to apply the required global index mappings; videos were copied byte-for-byte.

Metadata normalization

Two source-metadata inconsistencies were normalized in an immutable staging tree; the downloaded source folders were not modified.

1. In 28 folders, fps was attached to the five index features timestamp, frame_index, episode_index, index, and task_index. Staging removed fps only from those five non-temporal-schema fields. Camera, action, state, reward, done, and annotation metadata were retained.
2. In 28 folders, tasks.parquet contained one trailing description row that was never referenced by any episode/frame after slicing. Staging retained exactly the description indices referenced by the sliced Parquet data.

The source task tables contain 4,643 rows in total, while 4,615 rows are actually referenced by the sliced data. Global text deduplication then yields 4,610 output descriptions.

The merge also repairs the custom annotation mapping:

• annotation.human.task_description == task_index for every output frame
• annotation.human.task_name == source_task_index for every output frame

Index mapping

• index: contiguous 0..27,815,380
• episode_index: contiguous 0..29,999
• task_index: global description index 0..4,609
• annotation.human.task_description: equal to global task_index
• annotation.human.task_name: source task-folder index 0..299

meta/source_task_folders.parquet records the source/output episode and frame ranges for every folder. meta/provenance/task_folder_manifest.parquet adds the source descriptions and their global description indices.

Integrity verification

The completed merge was checked against every source folder, not by sampling.

• All 300 Parquet task files were compared.
• All 27,815,381 rows were checked.
• Unchanged payload columns were value-bit-exact: observation.state, action, next.reward, next.done, timestamp, and frame_index.
• Rewritten index/annotation columns were checked row-by-row against their deterministic source-to-output mapping.
• All 954 MP4 files were compared source-to-output with SHA-256.
• All 51,480,763,654 video bytes are covered by the SHA-256 manifest.

Verification artifacts:

• meta/provenance/integrity_report.json
• meta/provenance/parquet_bitcheck.parquet
• meta/provenance/video_sha256_manifest.parquet
• meta/provenance/task_folder_manifest.parquet
• meta/provenance/source_manifest.json
• meta/provenance/preflight_summary.json
• meta/provenance/merge_runtime.json

The SHA-256 of video_sha256_manifest.parquet is 6b10ccb06e767ea7fc33f562c7a911b8c5815172b73f3eba1bff913888838fc8.

Loading with LeRobot 0.6.1

from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    "Whalswp/robocasa_pretrain_human300_v3",
    root=Path.home() / "robocasa_pretrain_human300_v3",
)
print(len(dataset))  # 27,815,381

For training configs, use:

dataset:
  repo_id: Whalswp/robocasa_pretrain_human300_v3
  root: ${HOME}/robocasa_pretrain_human300_v3

Limitations

• This is the deterministic seed-0 preslice, not all episodes in the parent collection.
• Folder identity and language-description identity are different concepts; use annotation.human.task_name or meta/source_task_folders.parquet for the 300 folder-level tasks.
• Consult the upstream RoboCasa release and source repository for applicable asset, code, and dataset terms.

Citation

@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}
}

@inproceedings{robocasa2024,
  title={RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots},
  author={Soroush Nasiriany and Abhiram Maddukuri and Lance Zhang and Adeet Parikh and Aaron Lo and Abhishek Joshi and Ajay Mandlekar and Yuke Zhu},
  booktitle={Robotics: Science and Systems (RSS)},
  year={2024}
}