Galaxea Open-World Dataset

A roughly 500-hour dataset Galaxea collected at real-world locations using a single model of mobile bimanual robot.

The Galaxea Open-World Dataset is a real-robot dataset the embodied-AI company Galaxea (星海图) open-sourced in August 2025 alongside its G0 model paper. All the data was collected using the company's own R1 Lite mobile bimanual robot (23 degrees of freedom) at 11 real-world locations — homes, restaurants, retail stores, and offices — totaling about 100,000 trajectories, 150 task categories, 50 scenes, more than 1,600 object types, and 58 manipulation skills, adding up to roughly 500 hours. Each trajectory is split into atomic subtasks, with annotators choosing wording from standardized descriptions, and every subtask is labeled in both Chinese and English. The data is in LeRobot format, includes 4 camera streams (head plus both wrists), joint states, and IMU data, and can be downloaded from Hugging Face and the ModelScope community. Its distinguishing feature is using a single robot embodiment in real environments; the paper found that pretraining on this single-embodiment data was critical to the G0 model's performance.

ExampleIn the “tidy up fruit” task, the R1 Lite arranges fruit in a real environment; the whole trajectory is split into several subtasks, each labeled with a description in both Chinese and English.

Related

Galaxea AI · Galaxea R1 · Galaxea G0 Dual-System VLA · Real-Robot Data · Subtask Segmentation · LeRobotDataset

Sources

Galaxea Open-World Dataset and G0 Dual-System VLA Model (arXiv)
Galaxea Open-World Dataset (Hugging Face)
GalaxeaVLA 项目主页 (Chinese)

As of

2025-08