BridgeData V2
BridgeData V2 is a large-scale real-robot manipulation dataset collected at UC Berkeley RAIL Lab using the WidowX 250s robotic arm across diverse tabletop environments. Released in 2023 under the MIT license for commercial use, it contains 60,096 demonstration trajectories covering a wide range of household and kitchen manipulation tasks including pick-and-place, stacking, pushing, and tool use. The dataset spans multiple environments, camera configurations, and object sets, with demonstrations collected by multiple operators over two years. Data is distributed in HDF5 and Zarr formats on Hugging Face and Google Cloud Storage. BridgeData V2 is widely used as a pretraining dataset for generalised robot manipulation policies and is one of the primary datasets in the Open X-Embodiment collection. The WidowX 250s is a low-cost 6-DoF robotic arm widely used in academic manipulation research.
| Year | 2023 |
|---|---|
| Episodes | 60,096 |
| Embodiments | WidowX 250s |
| Modalities | rgb, proprioception |
| View type | third-person, wrist-cam |
| Task categories | manipulation, pick-and-place, cooking, cleaning |
| Data format | hdf5, zarr, mp4 |
| License | MIT |
| Access | open — commercial use permitted |
| Maintainer | UC Berkeley RAIL |
| Origin country | US |
What is it?
BridgeData V2 is a large-scale robot manipulation dataset collected at UC Berkeley RAIL Lab using the WidowX 250s robotic arm across diverse tabletop environments. Released in 2023 under MIT license, it contains 60,096 demonstration trajectories covering household and kitchen manipulation tasks. Demonstrations were collected across multiple environments by multiple operators over two years, with data distributed in HDF5 and Zarr formats on Hugging Face. BridgeData V2 is one of the primary constituent datasets of Open X-Embodiment and a standard pretraining dataset for generalised manipulation policies.
Who is it for?
BridgeData V2 targets researchers working on generalised manipulation policy training, particularly for low-cost 6-DoF arm platforms. The WidowX 250s is a widely accessible academic robot, making the dataset relevant for teams with limited hardware budgets. It is particularly useful as a pretraining source for researchers working on kitchen and household manipulation generalisation.
Key specifications
- Episodes: 60,096 demonstrations
- Embodiment: WidowX 250s (6-DoF arm)
- Tasks: Pick-and-place, stacking, pushing, tool use, cooking, cleaning
- Format: HDF5, Zarr, MP4
- License: MIT — commercial use permitted
- Access: Open — Hugging Face
How it compares
BridgeData V2 is one of the largest single-embodiment manipulation datasets for the WidowX arm. Compared to DROID (76,000 episodes, Franka Panda), BridgeData V2 covers more diverse environments but uses a lower-cost arm. It is complementary to DROID rather than competing — the two datasets together provide breadth across arm types and deployment contexts.
Limitations and access notes
BridgeData V2 covers tabletop manipulation only — no mobile base, legged locomotion, or humanoid data. The WidowX 250s has limited payload and reach compared to industrial arms like Franka Panda or UR5. MIT license permits unrestricted commercial use.
Linked professions
Frequently asked questions
What robot is BridgeData V2 collected on?
BridgeData V2 is collected using the WidowX 250s, a low-cost 6-DoF robotic arm manufactured by Trossen Robotics. The WidowX 250s is widely used in academic manipulation research due to its affordability and open-source software support.
Can BridgeData V2 be used commercially?
Yes. BridgeData V2 is licensed under MIT, permitting unrestricted commercial use, modification, and redistribution.
How does BridgeData V2 relate to Open X-Embodiment?
BridgeData V2 is one of the primary constituent datasets within the Open X-Embodiment collection. When you download Open X-Embodiment, BridgeData V2 is included as a sub-dataset under the bridge_v2 key.
What tasks does BridgeData V2 cover?
BridgeData V2 covers pick-and-place, stacking, pushing, sweeping, and tool use tasks across kitchen and tabletop environments. Tasks include moving objects between containers, sorting items, and simple food preparation operations.
How do I access BridgeData V2?
BridgeData V2 is available on Hugging Face at huggingface.co/datasets/rail-berkeley/bridge_data_v2. The full dataset is also available via the project GitHub repository. No registration is required.