BridgeData V2: A Dataset for Robot Learning at Scale
arXiv: 2308.12952 | Venue: CoRL 2023
Authors: Homer Walke, Kevin Black, Abraham Lee, Moo Jin Kim, Max Du, Chongyi Zheng, Tony Zhao, Philippe Hansen-Estruch, Quan Vuong, Andre He, Vivek Myers, Kuan Fang, Chelsea Finn, Sergey Levine
Abstract
BridgeData V2 is a large, diverse dataset of robot manipulation behaviors designed to enable scalable robot learning research. Collected on a publicly available low-cost WidowX 250 robot arm, it contains 60,096 trajectories spanning 24 environments and 13 task categories. The dataset supports open-vocabulary, multi-task learning conditioned on goal images or natural language instructions.
Dataset Scale
- Trajectories: 60,096
- Environments: 24
- Task categories: 13 skill types (pick-and-place, pushing, sweeping, door/drawer manipulation, etc.)
- Robot: WidowX 250 6-DoF arm (low-cost, publicly purchasable)
- Data split: ~85% human teleoperated, ~15% scripted
Hardware & Teleoperation
- Robot: WidowX 250 6-DoF arm
- Teleoperation: VR controller (SpaceMouse-style)
- Environment: Kitchen and tabletop settings
- Conditioning: Goal images OR natural language instructions
Key Design Philosophy
- Low-cost, reproducible hardware: WidowX 250 is affordable (~$3k) and widely available; enables community-wide data contribution
- Skill diversity over scene diversity: 13 skill types across 24 environments (vs. DROID’s 86 tasks across 564 scenes)
- Scaling laws verified: Performance improves monotonically with data volume and model capacity
- Generalization target: Cross-environment, cross-domain, cross-institution generalization
Relationship to Open-X Embodiment
BridgeData V2 is a major contributor to the Open X-Embodiment dataset. Its WidowX platform data is one of the most represented robot types in Open-X.
Comparison with DROID
BridgeData V2 predates DROID and uses a simpler, cheaper hardware setup. DROID expands the in-the-wild collection philosophy but uses higher-end Franka Panda robots and VR headset teleoperation. BridgeData V2 remains widely used as a baseline and fine-tuning target for VLA models.