Description
GalaxeaVLA is the open-source code repository behind G0.5, Galaxea's vision-language-action (VLA) model for general-purpose robot manipulation. Instead of using a vision-language model only to read a scene and hand off to a separate action network, G0.5 keeps the VLM as the actor: one transformer decoder generates reasoning tokens and action tokens in the same autoregressive stream. The repository hosts the training code, pretrained checkpoints and deployment entrypoints needed to run the model on real robots or in simulation.
Read More
The model conditions on multi-view camera images, an embodiment identifier, a natural-language instruction and the robot's proprioceptive state. It can first generate a short chain of embodied reasoning (subtask, bounding box, 2D gripper trace, motion hint) and then continue into structured action tokens in the same sequence, which are decoded into continuous controls and executed in chunks before the next observation is folded back in. A cross-embodiment action tokenizer maps different robot arms and mobile bases into a shared 27-dimensional action space, so one checkpoint can drive different embodiments. G0.5 also keeps several seconds of visual history through a spatio-temporal attention mechanism so it can react to recent motion, not just the current frame.
Galaxea publishes ready-to-use checkpoints on Hugging Face for zero-shot deployment on its own R1 Lite and R1 Pro humanoid robots, as well as for SO-100/101 arms, the DROID Franka setup, and the LIBERO and RoboTwin 2.0 simulation benchmarks. The repository includes fine-tuning scripts built on the g05-base checkpoint, a policy-server/client architecture for running inference over WebSocket on separate GPU and robot machines, and a companion Galaxea Open-World Dataset of real-world mobile-manipulation demonstrations in RLDS and LeRobot format.
It is aimed at robotics researchers and engineers who want to fine-tune or deploy a general-purpose manipulation policy on their own hardware rather than build a VLA model from scratch. Running the pretrained models needs an NVIDIA GPU (at least 8GB for inference, 70GB or more for full fine-tuning), a Linux environment with CUDA 12.8 and PyTorch 2.7.1, and native CUDA extensions such as flash-attn. The current release supersedes the earlier G0/G0Plus code line, which the maintainers keep accessible only through an older commit for reference.
- CategoryRobotics & AI Hardware
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsLinux
GalaxeaVLA's Core Features
Unified autoregressive VLA architecture that generates reasoning and action tokens in one transformer decoder stream
Cross-embodiment ActionCodec mapping arms, grippers and mobile bases into a shared 27-dimensional action space
Native chain-of-thought fields (Subtask, BBox, Trace, ActionHint) for task decomposition and object grounding
Multi-second visual memory via factorized spatial-temporal attention over camera history
Pretrained g05-base checkpoint plus task-specific checkpoints for R1 Lite, R1 Pro, SO-100/101, DROID, LIBERO and RoboTwin 2.0
Server/client deployment architecture: a policy server on a GPU machine and a robot-side client over WebSocket and msgpack
Fine-tuning scripts with multi-GPU distributed data-parallel support via torchrun
Companion Galaxea Open-World Dataset (500+ hours of real-world mobile manipulation data) in RLDS and LeRobot formats
Getting Started with GalaxeaVLA
Install: clone the repository and set up the environment with uv sync, then activate the virtual environment
Download checkpoints: pull the G05 model files from Hugging Face into a local checkpoints/ folder
Start the policy server: run scripts/serve_policy.py with the chosen checkpoint on a CUDA GPU
Connect the robot client: set the server address in the target embodiment's config.toml and launch its client (R1 Lite, R1 Pro, SO-100/101 or DROID)
Fine-tune (optional): run scripts/run/finetune.sh against the g05-base checkpoint to adapt the model to a new task or robot
GalaxeaVLA's Use Cases
- Humanoid manipulation research
- Low-cost arm deployment
- Simulation benchmarking
- Custom robot fine-tuning
