Generated illustration for BridgeData V2: A Dataset for Robot Learning at Scale — Paper 23 of 41Seedream 4.0 · GMI Studio

Illustration brief the model wrote

A robotic arm reaching into a cluttered kitchen sink with dishes, a toy bear on a counter, and a fruit bowl on a table; the gripper open over a mug; multiple cameras and varied backgrounds suggesting different environments.

Provenance. Brief written by deepseek-ai/DeepSeek-V4-Flash via GMI Cloud on 2026-09-18 from the paper's title, primary link and the editor's notes; reviewed editorially. Graphic rendered by seedream-4-0-250828 through the GMI Studio request queue (request d1ad34b4…), converted to PNG locally. Summaries can be wrong; the linked paper is the source of truth.

Atlas/Berkeley RAIL·Research lab·2023

BridgeData V2: A Dataset for Robot Learning at Scale

Model tl;dr

Introduces BridgeData V2, an open dataset of 60k multi-task manipulation trajectories across diverse environments for scalable robot learning.

datasetrobot learningmanipulationimitation learninggeneralizationbenchmarking

What changed

Model

Before BridgeData V2, robot learning datasets were often limited in scale, task variety, or environment diversity, hindering generalization. This dataset provided a large, standardized collection of 60k trajectories spanning kitchens, toys, and sinks, enabling pretraining of generalist policies and systematic benchmarking of manipulation algorithms.

So what, in plain English

Model

Think of it as a giant cookbook with hundreds of different recipes, kitchens, and utensils. A robot can practice from this book and learn to cook in any kitchen, not just one, because it has seen so many variations.

Contribution to industry

Model

BridgeData V2 became one of the most widely used open datasets for pretraining robot manipulation policies, adopted by academic labs and robotics startups. It enabled more robust generalization without requiring new data collection for each task, and set a benchmark for evaluating cross-environment transfer in manipulation.

Typed judgments

Jev

★★★★☆ manipulation (also robots) · Physical AI 97%

Real hardware 38% · Open code / data 94% · Foundation-model based 14% · Survey / benchmark-only 3%

Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.

Editor's notes

Summary. An open dataset of roughly 60,000 multi-task, multi-environment manipulation trajectories spanning kitchens, toys, sinks and diverse objects and scenes.

Industry. Became one of the most-used open manipulation datasets for pretraining policies and for benchmarking generalization across environments.

Read the paper ↗Official publication archive ↗All 2 Berkeley RAIL papers →