BridgeData V2 Robot Manipulation Dataset
Provider：UC Berkeley RAIL Lab
BridgeData V2 has 60,096 real WidowX manipulation trajectories across 24 environments and 13 skills, each with a natural-language instruction, open under CC BY 4.0.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: real WidowX 250 teleoperation trajectories with RGB/depth, proprioception and natural-language instructions, directly usable for policy training; the assets are real-world video and action logs rather than simulation-loadable 3D models, hence L1 as a real behavior data layer.
- Scale
- 60,096 trajectories (50,365 teleoperated demonstrations plus 9,731 scripted-policy rollouts) across 24 environments and 13 skills (The RAIL project page and the 2023 paper both report 60,096 trajectories across 24 environments and 13 skills, split into 50,365 teleoperated and 9,731 scripted rollouts. Checked 2026-10-25.)
- License
- CC BY — The dataset is released under CC BY 4.0 with attribution required for commercial use; code is licensed separately, and both the project page and paper state CC BY 4.0.
- Access
- Open
- Formats
- other
- Runtimes
- other
- Physics properties
- Collision geometryJoint structureSensor attributes
- Domain
- Other
- Published
- 2026-10-25
- Updated
- 2026-10-25
Why it is listed
BridgeData V2 is the reference real-world corpus for low-cost manipulation: collected on a roughly four-thousand-dollar WidowX 250 arm across twenty-four kitchen-like environments, it is deliberately cheap enough that other labs can reproduce the setup and compare results.
Readiness assessment
Every trajectory carries a crowdsourced natural-language instruction rather than a fixed task ID, so policies can be conditioned on free-form text at inference time. Data covers thirteen skills — pick-and-place, pushing, sweeping, stacking, cloth folding, drawer and door manipulation — recorded at 5 Hz with an average trajectory of about 38 timesteps.
Practical notes
The dataset is fully real, with no synthetic episodes, and is distributed as JPEG frame archives plus model training code and pre-trained weights; it is CC BY 4.0, so commercial reuse is allowed with attribution.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.