Robots, Data, and Foundation Models: Galaxea CTO Hang Zhao at CORL 2025 Sponsor Talk

In this talk

1. 0:25

   Building a robotics ecosystem

   Zhao says Galaxea combines robot hardware, pretrained models, and tools for evaluation, fine-tuning, and deployment. Its R1 Pro is designed for navigation and complex manipulation, while the smaller R1 Lite uses the same software stack. He says the company works with research labs, enterprises, and developers, and has more than 100 customers worldwide.

2. 3:44

   Open-world data and the G0 model

   Galaxea released an open-world dataset collected in apartments, hotels, offices, and warehouses, along with its G0 dual-system vision-language-action model. Zhao says the dataset had been downloaded more than 120,000 times and ranked first on Hugging Face and ModelScope. He cites tasks such as bed-making and says some downstream tasks used only 20 trajectories.

3. 6:41

   How G0 is trained and runs

   G0 pairs a slower vision-language system for understanding and planning with a faster system for motion and closed-loop control. Zhao gives rates of 2 hertz for reasoning and 20 hertz for control. Training proceeds through cross-embodiment pretraining, pretraining on Galaxea’s dataset for a single embodiment, and post-training with 20 to 100 task trajectories.

4. 7:57

   Data realism and evaluation

   Zhao argues that data from tidy collection sites may not prepare robots for unpredictable homes, and calls for messy environments and longer tasks. He says robotics evaluation depends on system details, since small changes in hardware or surroundings can shift results. His practices include replaying logs to check setup, establishing open-loop performance, and comparing multiple models in parallel.

5. 10:44

   Planned open-source releases

   Zhao previews releases planned for October 2025: 500 additional hours of R1 Pro data, which would bring the collection to 1,000 hours, with more annotations. Galaxea also planned G0 models in several sizes, a G0 Plus model for zero-shot pick-and-place, and tools for cross-robot calibration and inference.