OpenGalaxea/Galaxea-Open-World-Dataset
Preview • Updated • 25.6k • 58
G0.5 is Galaxea's pretrained autoregressive vision-language-action model for general-purpose robot control. It conditions on multi-view RGB observations, an embodiment identifier, natural-language task instruction, and robot proprioceptive state, then generates reasoning tokens and action tokens in one unified autoregressive stream.
In this repo, you can find:
g05-base/checkpoints/model_state_dict.pt: G0.5 base checkpoint.g05-libero/model.pt: G0.5 LIBERO checkpoint.g05-droid/checkpoints/model_state_dict.pt: G0.5 DROID checkpoint.g05-so101/checkpoints/model_state_dict.pt: G0.5 SO-100/101 checkpoint.g05-robotwin20/checkpoints/model_state_dict.pt: G0.5 RoboTwin 2.0 checkpoint. action_tokenizer.pt files: ActionCodec tokenizer weights for the corresponding checkpoints.dataset_stats.json files: normalization statistics for the corresponding checkpoints when provided.For code, configs, deployment scripts, and fine-tuning examples, please see the project repository:
https://github.com/OpenGalaxea/GalaxeaVLA
This repository contains materials released under different licenses depending on the commit date:
If you use our dataset or models, please cite:
@article{galaxea2026g05,
title={Galaxea G0.5 Technical Report},
author={Galaxea Team},
year={2026},
url={https://opengalaxea.github.io/G05/Galaxea_G0_5.pdf}
}