BridgeData V2: A Dataset for Robot Learning at Scale

Walke, Black, Lee, Kim, Du, Zheng, Zhao, Hansen-Estruch, Vuong, et al. · 2023 · arXiv preprint · arXiv:2308.12952 · PDF

Dongyu supplement. Added as a candidate missing/adjacent dataset paper after searching primary arXiv sources. This is a draft summary for triage, not a full paper read.

One-liner. An open robot manipulation dataset with 60,096 trajectories across 24 environments, compatible with natural-language and goal-image conditioning.

Setup

• Datasets / benchmarks: Introduces BridgeData V2: 60,096 trajectories across 24 environments on a publicly available low-cost robot. The authors publicly share the dataset and pretrained models.
• Hardware / simulator: Publicly available low-cost robot; broad environment and task variability.

Method

Dataset plus experiments across imitation learning and offline RL methods; supports open-vocabulary multi-task learning with goal images or natural language instructions.

Why it matters for the map

Good anchor for open language-conditioned robot learning data, even though it is not centered on non-visual sensing.

Limitations / open questions

Vision/language/action focused; little direct tactile/audio/force/thermal coverage.

Source note

arXiv lines 38-40 report scale, language compatibility, and public sharing.