README.md

---
pretty_name: Bridge LeRobot v3
library_name: lerobot
tags:
- robotics
- robot-learning
- vision-language-action
- bridgedata
- bridge
- lerobot
- LeRobotDataset-v3
license: openmdw-1.0
---

# BridgeData2 LeRobot v3

## Dataset Summary

`nvidia/bridge_lerobot_v3` is a LeRobotDataset v3.0 conversion of the BridgeDataset / BridgeData V2 robot manipulation dataset. BridgeData V2 is a large-scale real-world robotics dataset collected to support scalable robot learning, including imitation learning, offline reinforcement learning, and open-vocabulary multi-task policies conditioned on goal images or natural-language instructions.

This repository packages Bridge trajectories in the LeRobot v3 layout with Parquet state/action data, MP4 video observations, and structured LeRobot metadata. The converted repository's `meta/info.json` reports 50,415 episodes, 1,801,090 frames, 22,199 task IDs, and a 5 Hz sampling rate.

## Dataset Details

### Dataset Description

BridgeData V2 contains robot manipulation behaviors collected across varied tabletop and toy-kitchen environments. The upstream dataset was designed to study generalization across tasks, objects, environments, institutions, and conditioning modes such as language instructions and goal images.

This LeRobot v3 conversion preserves the dataset for workflows that use the LeRobot data loader and Hugging Face Hub-native robotics dataset conventions. The conversion changes the storage layout: frame-level tabular data is stored in Apache Parquet shards, visual observations are stored as MP4 videos, and metadata records the schema, feature statistics, task IDs, episode boundaries, and path templates.

This dataset is ready for commercial or non-commercial uses.

## Dataset Owner(s)

NVIDIA Corporation

## Dataset Creation Date

2026-05-26

## Version:
v1.0 <br>

## License/Terms of Use

This dataset is released under the [OpenMDW1.1](https://openmdw.ai/)

### Dataset Sources

- Repository: https://huggingface.co/datasets/nvidia/bridge_lerobot_v3
- BridgeData V2 project page: https://rail-berkeley.github.io/bridgedata/
- BridgeData V2 paper: https://arxiv.org/abs/2308.12952
- LeRobotDataset v3.0 documentation: https://huggingface.co/docs/lerobot/lerobot-dataset-v3

## Intended Use

This dataset is intended for research and development in:

- robot imitation learning
- offline reinforcement learning
- vision-language-action model training
- goal-conditioned and language-conditioned robot policy learning
- multi-camera visuomotor policy learning
- evaluation of LeRobot-compatible training, streaming, and data-loading workflows

### Out-of-Scope Use

This dataset is not intended to be used as the sole validation source for safety-critical robot deployment. Policies trained on this data should be evaluated in the target environment, with appropriate robot safety controls, before any physical deployment.

## Dataset Characterization
**Data Collection Method**

[Not Applicable]

**Labeling Method**

[Not Applicable]

## Dataset Format

The repository follows the LeRobotDataset v3.0 layout:

```text
.
+-- data/
|   `-- chunk-000/
|       `-- file-*.parquet
+-- meta/
|   +-- info.json
|   +-- stats.json
|   +-- tasks.parquet
|   `-- episodes/
|       `-- chunk-000/
|           `-- file-*.parquet
`-- videos/
    `-- observation.images.<camera_key>/
        `-- chunk-000/
            `-- file-*.mp4
```

### Dataset Quantification

The following values are taken from the repository's `meta/info.json`.

| Field | Value |
|---|---:|
| LeRobot codebase version | `v3.0` |
| Episodes | 50,415 |
| Frames | 1,801,090 |
| Task IDs | 22,199 |
| Split | `train: 0:50415` |
| FPS | 5 |
| Average episode length | 35.7 frames, about 7.1 seconds |
| Robot type in metadata | `null` |

The upstream BridgeData V2 project describes the source data as collected on a WidowX 250 6DOF robot arm at 5 Hz. The converted repository does not populate `robot_type` in `meta/info.json`.

Total Data Storage: 90 GB

### Features

The following feature schema is declared in `meta/info.json`.

| Feature | Type | Shape / Details |
|---|---|---|
| `observation.images.image_0` | video | RGB, 480 x 640, AV1 MP4, 5 FPS, no audio |
| `observation.images.image_1` | video | RGB, 480 x 640, AV1 MP4, 5 FPS, no audio |
| `observation.images.image_2` | video | RGB, 480 x 640, AV1 MP4, 5 FPS, no audio |
| `observation.images.image_3` | video | RGB, 480 x 640, AV1 MP4, 5 FPS, no audio |
| `observation.state` | `float32` | 7-dimensional robot state |
| `action` | `float32` | 7-dimensional robot action |
| `timestamp` | `float32` | Frame timestamp |
| `frame_index` | `int64` | Frame index within episode |
| `episode_index` | `int64` | Episode index |
| `task_index` | `int64` | Task ID |
| `index` | `int64` | Global frame index |

### File Format

- Frame-level state/action data: Apache Parquet under `data/`.
- Episode metadata: chunked Parquet under `meta/episodes/`.
- Task metadata: `meta/tasks.parquet`.
- Dataset schema and statistics: `meta/info.json` and `meta/stats.json`.
- Video observations: AV1-encoded MP4 files under `videos/`.

## Loading

Authenticate with Hugging Face if required, then download or stream the dataset using LeRobot-compatible tooling.

```python
from huggingface_hub import snapshot_download

repo_dir = snapshot_download(
    repo_id="nvidia/bridge_lerobot_v3",
    repo_type="dataset",
    token=True,
)

print(repo_dir)
```

Example LeRobot usage:

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("nvidia/bridge_lerobot_v3")
print(dataset.num_episodes)
print(dataset.meta.info["features"].keys())
```

## Dataset Creation

### Source Data

The source dataset is BridgeData V2, a real-world robot manipulation dataset with diverse tasks, objects, camera poses, and environments. The BridgeData V2 project page reports 60,096 trajectories across 24 environments and 13 skills, including teleoperated demonstrations and scripted pick-and-place rollouts.

This converted repository reports 50,415 episodes in `meta/info.json`; users should treat the repository metadata as authoritative for this LeRobot package and consult upstream BridgeData V2 documentation for the full source dataset description.

### Conversion

This repository converts Bridge data into LeRobotDataset v3.0. In v3, multiple episodes can be concatenated into larger Parquet and MP4 shard files, while metadata is used to recover per-episode boundaries and feature statistics. This layout reduces file-system pressure and supports Hub-native loading and streaming workflows.

## Risks, and Limitations

- The dataset reflects the embodiment, control frequency, camera setup, task distribution, and environments of the source BridgeData V2 collection.
- The upstream data is concentrated around tabletop and toy-kitchen manipulation tasks; policies trained only on this dataset may not generalize to other robots, objects, lighting, homes, labs, or industrial settings.
- Natural-language task annotations are inherited from the source/conversion pipeline and should be inspected before task-specific filtering or evaluation.
- The repository metadata reports a converted package with fewer episodes than the full upstream BridgeData V2 trajectory count.
- This dataset should not be used as the sole basis for validating safe real-world robot behavior.

## Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).

## Citation

If you use this dataset, please cite the original BridgeData V2 paper:

```bibtex
@inproceedings{walke2023bridgedata,
    title={BridgeData V2: A Dataset for Robot Learning at Scale},
    author={Walke, Homer and Black, Kevin and Lee, Abraham and Kim, Moo Jin and Du, Max and Zheng, Chongyi and Zhao, Tony and Hansen-Estruch, Philippe and Vuong, Quan and He, Andre and Myers, Vivek and Fang, Kuan and Finn, Chelsea and Levine, Sergey},
    booktitle={Conference on Robot Learning (CoRL)},
    year={2023}
}
```

Please also cite or reference LeRobot if you use LeRobot tooling, dataset loaders, or streaming interfaces.

## References

- BridgeData V2 project page: https://rail-berkeley.github.io/bridgedata/
- BridgeData V2 arXiv paper: https://arxiv.org/abs/2308.12952
- BridgeData V2 Hugging Face paper page: https://huggingface.co/papers/2308.12952
- LeRobotDataset v3.0 blog: https://huggingface.co/blog/lerobot-datasets-v3
- LeRobotDataset v3.0 documentation: https://huggingface.co/docs/lerobot/lerobot-dataset-v3