GT-111/intern-a1-v3-ee

InternData-A1 (LeRobot v3 EE)

This is a LeRobot v3 format conversion of the InternData-A1 dataset, reorganized into a canonical 16D end-effector (EE) pose representation for multi-dataset robot learning.

Original Dataset

InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy

Yang Tian, Yuyin Yang, Yiman Xie, Zetao Cai, Xu Shi, Ning Gao, Hangxu Liu, Xuekun Jiang, Zherui Qiu, Feng Yuan, Yaping Li, Ping Wang, Junhao Cai, Jia Zeng, Hao Dong, Jiangmiao Pang. InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy. arXiv:2511.16651, 2025.

• Original Source: https://huggingface.co/datasets/InternRobotics/InternData-A1
• arXiv: https://arxiv.org/abs/2511.16651

InternData-A1 is a hybrid synthetic-real manipulation dataset containing over 630k trajectories and 7,433 hours across 4 embodiments (Franka, Genie1, Lift2, Split-ALOHA), 18 skills, 70 tasks, and 227 scenes, covering rigid, articulated, deformable, and fluid-object manipulation.

License

This dataset is released under CC BY-NC-SA 4.0, consistent with the original InternData-A1 license.

Important: By downloading this dataset, you agree to the InternData-A1 Community License Agreement. The original dataset is gated — please also request access at the original Hugging Face repo.

Conversion Details

What we changed

1. Multi-Embodiment Canonicalization: The original dataset spans 4 robot embodiments (Franka, Genie1, Lift2, Split-ALOHA) with varying proprioception layouts. We unified all embodiments into a common 16D bimanual EE pose representation:

   • [left_x, left_y, left_z, left_qw, left_qx, left_qy, left_qz, left_gripper, right_x, right_y, right_z, right_qw, right_qx, right_qy, right_qz, right_gripper]
   • For single-arm embodiments (Franka, Genie1), the right-arm half is filled with learned padding values.
   • Gripper values normalized to [0, 1].

2. Forward Kinematics: Where raw data stores joint positions, we compute forward kinematics to derive EE poses (position + quaternion).

3. Canonical Camera Keys: Unified camera naming across embodiments:

   • observation.images.cam_high — head/top camera
   • observation.images.cam_left_wrist — left wrist camera
   • observation.images.cam_right_wrist — right wrist camera (when available)

4. Image Masks: Per-frame boolean masks indicate which camera views are valid (some embodiments lack certain cameras).

5. LeRobot v3 Format: Converted to the latest LeRobot v3 dataset layout with sharded video storage (MP4) and Parquet-based frame data.

What we preserved

• All available camera views per embodiment
• Original episode structure and task labels
• Frame-level timestamps
• Embodiment metadata (openpi_embodiment in info.json)

Child Datasets

Embodiment	Episodes	Frames	Type
franka	96,930	66,441,068	Single-arm (sim+real)
genie1	14,064	7,713,365	Single-arm (sim)
lift2	205,877	153,366,384	Bimanual (sim)
split_aloha	170,877	150,163,119	Bimanual (sim)

Dataset Structure

lerobot_v3_ee/
├── franka/
│   ├── data/chunk-*/file-*.parquet
│   ├── videos/observation.images.cam_high/chunk-*/file-*.mp4
│   └── meta/{info.json, tasks.json, episodes.jsonl, stats.json}
├── genie1/
│   └── ...
├── lift2/
│   └── ...
├── split_aloha/
│   └── ...
└── build_summary.json

Usage

from lerobot.datasets import LeRobotDataset

# Load a specific embodiment
dataset = LeRobotDataset(
    repo_id="GT-111/intern-a1-v3-ee",
    root="franka",  # or genie1, lift2, split_aloha
)

For training with the LWM-VLA / OpenPI framework (automatically discovers all child datasets):

from openpi.training.config import MultiDatasetPretrainDatasetSpec

MultiDatasetPretrainDatasetSpec(
    repo_id="GT-111/intern-a1-v3-ee",
    dataset_type="intern_a1",
    weight=0.2,
)

Citation

If you use this dataset, please cite both the original InternData-A1 paper and this conversion:

@article{tian2025interndata,
  title={InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy},
  author={Tian, Yang and Yang, Yuyin and Xie, Yiman and Cai, Zetao and Shi, Xu and Gao, Ning and Liu, Hangxu and Jiang, Xuekun and Qiu, Zherui and Yuan, Feng and Li, Yaping and Wang, Ping and Cai, Junhao and Zeng, Jia and Dong, Hao and Pang, Jiangmiao},
  journal={arXiv preprint arXiv:2511.16651},
  year={2025}
}

Version History

• v3.0 (current): LeRobot v3 conversion with canonical 16D EE pose layout across 4 embodiments