Dataset
BC-Z
Google large-scale real robot demos with language-conditioned multi-task manipulation.
DatasetFeatured
BridgeData V2
Large household WidowX dataset (~60K trajectories) with rich language conditioning.
Why: Low hardware bar and strong language labels for real-data starters
Overview
BridgeData V2 is a large real-robot multi-domain demonstration dataset widely used for finetuning open VLAs and imitation policies. It improves coverage over BridgeData V1 and appears in many OXE mixtures.
Watch camera extrinsics, action definitions, and language annotations when converting into your trainer’s format.
Learning Path: primary real finetune dataset next to DROID.
Who it is for
Low-cost arm labs; language-conditioned IL/VLA finetunes.
Key highlights
- ~60K real trajectories
- Strong language annotations
- Widely used in open VLA papers
When to use
- WidowX finetune
- Language-conditioned household tasks
When not to use
- Bimanual or humanoid-first research without retargeting
Getting started
- 1Fetch dataset via official instructions
- 2Filter by task keywords
- 3Finetune ACT/OpenVLA-style policies
Papers & reading
How it compares
Caveats & pitfalls
- Embodiment gap if your arm kinematics differ a lot.
Content reviewed 2026-07-28