BridgeData V2: A Dataset for Robot Learning at Scale
Source
- Raw Markdown: paper_bridge-data-v2-2023.md
- PDF: paper_bridge-data-v2-2023.pdf
- Preprint: arXiv 2308.12952
Core Claim
BridgeData V2 is a real-robot manipulation dataset used as an important substrate for language-conditioned robot policies and robotic world-model evaluation.
Sensor-Time-Series Notes
- The useful modeling unit is a language-conditioned manipulation trajectory with image observations and robot control inputs.
- The dataset is especially relevant to action-conditioned latent world models because several later studies use Bridge-style rollouts to test whether generated future observations preserve action-relevant state.
- It should be treated as visual robot trajectory data rather than as a generic numeric forecasting benchmark.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Action-conditioned trajectories | adjacent | The paper describes language- or goal-conditioned manipulation trajectories with observations and robot actions across many environments. | Robotics visual-control data, not numeric operational time series or a digital-world intervention benchmark. |
| Context interface | adjacent | Tasks can be conditioned by natural language instructions or goal images, forcing policies to use task context rather than infer the task only from initial state. | Needs a typed context/action schema transferable to telemetry, business, or cyber-physical systems. |
Links Into The Wiki
- Foundation Time-Series Model Research Agenda
- Robotics Time-Series Modeling
- Action-Conditioned Time-Series Datasets
- World Models
Open Questions
- Which representation of BridgeData V2 actions is most useful for cross-dataset training: raw robot commands, normalized end-effector deltas, or learned action tokens?
- How much proprioceptive or force/contact information is needed in addition to camera history for robust Bridge-style world models?