We are gradually open-sourcing the dataset and model. Progress will be updated here:

• Aug 23, 2025

  • Release Galaxea Open-World Dataset.
  • Now our Open-Galaxea-Dataset is available at Huggingface and Modelscope!

• Sep 9, 2025

  • Release G0-VLA pretrained model weights.
  • Now our pretained weight is available at Huggingface and Modelscope!

• Sep 17, 2025

  • Release G0-VLA real-robot inference code.

• Sep 17, 2025

  • Release G0-VLA fine-tuning code.

• Oct 7, 2025

  • Now Lerobot Format Galaxea Open-World Dataset is available at Huggingface!

• Later in 2025

  • 🔮 More updates to come (extended datasets, improved models, additional tools).

We introduce Galaxea Open-World Dataset, a large-scale, high-quality robot behavior dataset collected in authentic human living and working environments.
We also present G0, a dual-system VLA model that combines:

• G0-VLM: a multimodal planner for high-level reasoning and subtask planning.
• G0-VLA: a real-time executor for precise low-level action control.

The dataset and model are designed to advance real-world, long-horizon, and few-shot robotic manipulation.

• 500+ hours of real-world mobile manipulation data.
• All data collected using one uniform robotic embodiment for consistency.
• Fine-grained subtask language annotations.
• Covers residential, kitchen, retail, and office settings.
• Dataset in RLDS format.

See more dataset (formats and examples) details here.

To run our pretrained models in this repository, you will need an NVIDIA GPU with at least the following specifications. These estimations assume a single GPU, but you can also use multiple GPUs with model parallelism to reduce per-GPU memory requirements by configuring --nnodes and--nproc-per-node in the fine-tune start shell script.

git clone https://github.com/OpenGalaxea/G0
conda env create -f glx_vla.yml
conda activate g0

# Install Pacakges from Code
git clone https://github.com/kvablack/dlimp
cd dlimp
pip install -e .

To fine-tune our model with your own data, you should follow three steps:

1. Convert your data to a RLDS dataset. You can follow data converter open-sourced by OpenVLA.

2. Defining training configs and running training:

   Defining Training Configs

   You can find a training config template at vla/config/r1_lite/r1lite_fine_tune_example.yml

   Running Training

   cd G0
   conda activate g0
   
   # For Single Nodes Post-Training
   torchrun \
       --standalone \
       --nnodes 1 \
       --nproc-per-node <num-gpus> \
       finetune.py --config <your-training-config-path>
   
   

3. Running Inference in with real world with Galaxea R1Lite Robot (ROS1)

   See detailed commands and launch methods here.

1. Inference: Support either BF16 or FP32. You can change data type by specifying dtype parameter while launching inference.
2. Training: Support either BF16 or FP32. You can enable BF16 by setting enable_bf16: True in the training config file. Our open-sourced pretrained weight is trained with BF16.

We will collect common issues and their solutions here. If you encounter an issue, please check here first. If you can't find a solution, please file an issue on the repo.

This project builds upon prior work from the open-source community. The implementation was inspired by open-pi-zero, OpenVLA, Octo, and Openpi, and the experiments make use of datasets including OXE, RDT, BridgeV2, and DROID. We sincerely thank the authors of these projects for making their code and data publicly available.

If you use our dataset or models, please cite:

@article{galaxea2025,
  title={Galaxea G0: Open-World Dataset and Dual-System VLA Model},
  author={Galaxea Team},
  journal={arXiv preprint arXiv:2509.00576v1},
  year={2025}
}