Galaxea Open-World Dataset & VLA Models

📢 News

[Jun 1, 2026] Introducing G0.5, our latest autoregressive VLA model with state-of-the-art performance Project Page! Model weights and code are coming soon.

[Feb 12, 2026] Update G0Plus pre-trained weights trained on larger-scale teleoperation and web data. Release G0Tiny (250M, SmolVLM2 backbone) for R1 Pro Orin edge deployment. New out-of-the-box demos: Fold Towels and Handover Gift (on-device G0Tiny inference via TensorRT at up to 10 Hz). Add openpi-based pi0/pi0fast fine-tuning support.

[Jan 4, 2026] We are releasing G0Plus, our latest pre-trained VLA model for multi-task robot manipulation.

[Oct 7, 2025] Now Lerobot Format Galaxea Open-World Dataset is available at Huggingface!

[Sep 17, 2025] Release G0-VLA fine-tuning and real-robot inference code.

[Sep 9, 2025] Release G0-VLA pretrained model weights. Huggingface and Modelscope!

[Sep 9, 2025] Release Galaxea Open-World Dataset. Huggingface and Modelscope!

📌 Overview

GalaxeaVLA is an open-source project dedicated to advancing real-world, long-horizon, and few-shot robot manipulation.

1. Galaxea Open-World Dataset

   • 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/LeRobot format.

2. Easy-to-Use Fine-Tuning Framework

   • Fully compatible with the LeRobot dataset format and scalable to large, real-world datasets.
   • Modular design enables easy extension and adaptation for new tasks and environments.

3. Model Checkpoints & An Out-of-the-Box Demo!

   • G0Plus_3B-base: A powerful pre-trained model with 2k hours+ real-world robot data for fine-tuning on custom tasks.
   • G0Tiny_250M-base: A lightweight pre-trained model with 1k hours of R1 Pro VR teleoperation data, with only 250M parameters for on-device deployment on the R1 Pro Orin platform.
   • G0Plus_3B-pick_and_place: A deployment-ready checkpoint, post-trained for robust pick-and-place performance in the wild.
   • Out-of-the-Box Pick Up Anything Demo: a Dockerfile and step-by-step guides for quick setup and reproducible experiments.
   • Out-of-the-Box Fold Towels Demo: a Dockerfile and step-by-step guides for quick setup and reproducible experiments.
   • Out-of-the-Box Handover Gift Demo: a step-by-step guide for on-device G0Tiny VLA inference on R1 Pro Orin.

🚀 Galaxea Open-World Dataset

Key features

• 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 and LeRobot format.

See more dataset (formats and examples) details here.

⚙️ GalaxeaVLA Getting Started

GPU Requirements

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.

Mode	Memory Required	Example GPU
Inference	> 8 GB	RTX 3090 / RTX 4090 (Recommended)
Fine-Tuning (Full)	> 70 GB	A100 (80GB) / H20 (96GB)

Installation

git clone https://github.com/OpenGalaxea/GalaxeaVLA
cd GalaxeaVLA
uv sync --index-strategy unsafe-best-match
source .venv/bin/activate

uv pip install -e .
uv pip install -e .[dev]

Note that before you run the installation:

1. Recommend to install uv without using a conda environment.
2. Recommend to add env variables at the beginning of your terminal, if you are in the country:

   export UV_DEFAULT_INDEX=https://mirrors.aliyun.com/pypi/simple/
   export UV_PYTHON_INSTALL_MIRROR=https://gh-proxy.com/https://github.com/astral-sh/python-build-standalone/releases/download
   

Model Checkpoints

Model	Use Case	Description	Checkpoint Path
G0_3B-base	Fine-Tuning	Base G0-VLA Model for fine-tuning	https://huggingface.co/OpenGalaxea/G0-VLA/blob/main/G0_3B_base.pt
G0Plus_3B-base	Fine-Tuning	Base G0Plus-VLA Model for fine-tuning	https://huggingface.co/OpenGalaxea/G0-VLA/tree/main/G0Plus_3B_base
G0Tiny_250M-base	Fine-Tuning	Lightweight G0Tiny-VLA Model (250M) for edge deployment on R1 Pro Orin	https://huggingface.co/OpenGalaxea/G0-VLA/tree/main/G0Tiny_260120
G0Plus_3B-pick_and_place	Deployment	Pick-and-Place Demo in the Wild	https://huggingface.co/OpenGalaxea/G0-VLA/tree/main/G0Plus_PP_CKPT

Inference on Real Robot

To run inference on a real Galaxea R1Lite robot using our pre-trained G0Plus model:

1. Make sure to finish the above installation steps first.

2. Then, follow steps and refer more details in our accompanying repo EFMNode.

🔥 Fine-Tuning Base Models on Galaxea Robots

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

1. Create your own task configs in configs/tasks/real/. You can adapt it from our configs demos: G0Plus on R1Lite or G0Tiny on R1Pro.

2. Install the required packages

   sudo apt install ffmpeg
   

3. Set your environment variables

   • HF_DATASETS_CACHE: An empty directory for HF-related caches.
   • GALAXEA_FM_OUTPUT_DIR: An empty directory for checkpoints and logs output.
   • GALAXEA_FM_DATASET_STATS_CACHE_DIR: A directory for caching dataset normalization statistics.
   • SWANLAB_API_KEY: Your SwanLab API key.

   export HF_ENDPOINT=https://hf-mirror.com
   export HF_DATASETS_CACHE=<YOUR_HF_CACHE_PATH>
   export GALAXEA_FM_OUTPUT_DIR=<YOUR_OUTPUT_DIR>
   export GALAXEA_FM_DATASET_STATS_CACHE_DIR=<YOUR_STATS_CACHE_DIR>
   export SWANLAB_API_KEY=<YOUR_SWANLAB_API_KEY>
   

4. Running fine-tuning

   bash scripts/run/finetune.sh <num_of_gpu> <task_path>
   
   # examples:
   bash scripts/run/finetune.sh 8 real/g0plus_r1lite_finetune
   bash scripts/run/finetune.sh 8 real/g0tiny_r1pro_finetune
   bash scripts/run/finetune.sh 8 real/pi0_r1lite_finetune
   bash scripts/run/finetune.sh 8 real/pi0fast_r1lite_finetune
   

FAQs of Fine-tuning

1. Q: How to convert my data to a LeRobot dataset?

   A: The demo datasets are provided on HuggingFace for easy trying.

2. Q: Cannot view the training logs in the SwanLab?

   A: Make sure you set your own swanlab workspace in train.yaml.

3. Q: Cannot find the pre-trained model?

   A: We use google/paligemma-3b-pt-224 and HuggingFaceTB/SmolVLM2-500M-Video-Instruct as the pre-trained models, you should modify them twice in g0plus.yaml and g0tiny the same as your actual paths (default: /To/Your/Path/google/paligemma-3b-pt-224 and /To/Your/Path/HuggingFaceTB/SmolVLM2-500M-Video-Instruct).

4. Q: Out of Memory (OOM) error?

   A: Make sure you have enough GPU memory as mentioned above. Or, reduce the batch_size in g0plus.yaml (default: 4).

🔥🔥 Out-of-the-Box Demos

1. Pick Up Anything
2. Fold Towels
3. Handover Gift

Feel free to raise an issue if you have any questions.

Acknowledgement

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.

📜 Citation

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}
}

License

This repository contains materials released under different licenses depending on the commit date:

• Apache-2.0 (Legacy): All content committed before 2026-01-04 is licensed under the Apache License 2.0.
• G0 PLUS Community License Agreement (Current): All content committed on or after 2026-01-04 is licensed under the G0 PLUS Community License (Non-Commercial + Limited Patent License). See G0 Plus Community License Agreement.

For avoidance of doubt, the licensing boundary is the first commit that introduced the G0 PLUS license switch:

• Boundary commit (first under G0 PLUS license): 318207fe6d994d0ecaf8f7d7ebb9b96fec5ebf56.

What you can do under the G0 Plus Community License

You may use, reproduce, modify, and distribute the G0 Plus materials only for non-commercial purposes, such as academic research, personal use, education, and evaluation. Commercial use (including production deployment, providing services to third parties, or productization) requires a separate commercial license from us.

Notices and attribution

If you redistribute any part of the G0 Plus materials, you must include:

• a copy/link of G0 Plus Community License Agreement, and
• the NOTICE file in this repository, and
• prominent notices on modified files indicating changes.

  Star History