GalaxeaVLA  by OpenGalaxea

Vision-Language-Action models for robot control

Created 11 months ago
681 stars

Top 49.0% on SourcePulse

GitHubView on GitHub
Project Summary

Summary OpenGalaxea/GalaxeaVLA presents G0.5, an autoregressive vision-language-action (VLA) model for general-purpose robot control. It addresses long-horizon task execution by unifying embodied reasoning and action generation within a single transformer decoder, targeting robotics researchers and engineers.

How It Works G0.5 utilizes a unified autoregressive transformer decoder to generate reasoning and action tokens sequentially. It conditions on multi-view RGB, embodiment ID, language instructions, and proprioceptive states. Key innovations include a Cross-embodiment ActionCodec mapping diverse actions to a shared 27D space via VQ-tokenization, and multi-second visual history integration through factorized spatial-temporal attention. This unified approach improves grounding and long-horizon task performance.

Quick Start & Requirements 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]

Hardware: NVIDIA GPU required; > 8 GB VRAM for inference (RTX 3090/4090 recommended), > 70 GB VRAM for full fine-tuning (A100/H20 recommended). Deployment guides and demos are available for R1 Lite, SO-100/101, and DROID.

Highlighted Details Achieves state-of-the-art performance: 82.5% zero-shot success on DROID, 98.9% on LIBERO, and 76.7% real-world fine-tuning success on R1-Lite/Pro. Utilizes the Galaxea Open-World Dataset (500+ hours, RLDS/LeRobot format). Offers multiple pre-trained checkpoints (G0.5-base, G0Tiny_250M-base, etc.).

Maintenance & Community Active development with recent updates (June 2026). No explicit community channels or roadmap details are provided in the README.

Licensing & Compatibility Content from June 16, 2026, onwards is under the G0.5 Community License (Non-Commercial + Limited Patent License). Earlier content may be Apache-2.0 or G0 PLUS Community License (also non-commercial). Commercial use requires a separate license. Redistribution requires including license copies and NOTICE text.

Limitations & Caveats The G0.5 Community License restricts commercial adoption. Fine-tuning requires high-end GPU hardware (>70GB VRAM), potentially limiting accessibility.

Health Check
Last Commit

1 week ago

Responsiveness

Inactive

Pull Requests (30d)
0
Issues (30d)
5
Star History
48 stars in the last 30 days

Explore Similar Projects

Feedback? Help us improve.