Galaxea DynamicsChina2026

G0.5

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

RobotsR1 Liteby Galaxea Dynamics AR1 Proby Galaxea Dynamics A

Key figures as stated by the developer

5 of 6 stated · 1 gap

Model type ⁠A vision-language-action embodied reasoning
Controls ⁠A arms and hands whole body
Inputs ⁠A camera language robot state
Access ⁠B open download
Parameters not stated
Base model ⁠A Qwen3.5 2B VLM

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (6)

Model type

Developer wording: "is Galaxea's pretrained autoregressive vision-language-action model for general-purpose robot control."; "The model can first emit embodied reasoning, then continue with structured action tokens in the same generation stream.".

Controls

README: "Action tokens are decoded into continuous robot controls, executed in chunks, and replanned closed-loop from new observations." "A learned residual vector-quantized action tokenizer represents semantically aligned motion groups such as left arm, right arm, and lower body." The paper: "the model acquires a factored whole-body control prior where each motion group can be independently predicted". "Lower-body actuation is represented in the unified action space, but is not evaluated separately in this work."

Inputs

README: "G0.5 conditions on multi-view RGB observations, an embodiment identifier, natural-language task instruction, and robot proprioceptive state."

Access

Hugging Face: "This repository is publicly accessible, but you have to accept the conditions to access its files and content". The weights are under a non-commercial licence.

Parameters

Note "G0.5 is initialized from a Qwen3.5 2B VLM and pretrained on robot demonstrations from 14 embodiments together with large-scale web and embodied VQA data."

Base model

README: "G0.5 is initialized from a Qwen3.5 2B VLM and pretrained on robot demonstrations from 14 embodiments together with large-scale web and embodied VQA data."

Full specifications

8 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action` `embodied reasoning`	A Developer wording: "is Galaxea's pretrained autoregressive vision-language-action model for general-purpose robot control."; "The model can first emit embodied reasoning, then continue with structured action tokens in the same generation stream.".
Inputs	`camera` `language` `robot state`	A README: "G0.5 conditions on multi-view RGB observations, an embodiment identifier, natural-language task instruction, and robot proprioceptive state."
Controls	`arms and hands` `whole body`	A README: "Action tokens are decoded into continuous robot controls, executed in chunks, and replanned closed-loop from new observations." "A learned residual vector-quantized action tokenizer represents semantically aligned motion groups such as left arm, right arm, and lower body." The paper: "the model acquires a factored whole-body control prior where each motion group can be independently predicted". "Lower-body actuation is represented in the unified action space, but is not evaluated separately in this work."
Degrees of freedom controlled	not stated	Note "Heterogeneous robot actions are mapped into a shared 27-dimensional action space:" It is a shared space across robots, not one robot's joint count. ⁠A
Size and speed
Parameters	not stated	Note "G0.5 is initialized from a Qwen3.5 2B VLM and pretrained on robot demonstrations from 14 embodiments together with large-scale web and embodied VQA data." ⁠A
Control rate	not stated
Where it runs and what it is built on
Runs where	not stated	Note README: "The GPU machine runs the G0.5 policy server from this repository, while the robot-side client collects raw observations, sends them over WebSocket + msgpack, receives action chunks, and executes them." ⁠A
Base model	Qwen3.5 2B VLM	A README: "G0.5 is initialized from a Qwen3.5 2B VLM and pretrained on robot demonstrations from 14 embodiments together with large-scale web and embodied VQA data."
Openness
Access	open download	B Hugging Face: "This repository is publicly accessible, but you have to accept the conditions to access its files and content". The weights are under a non-commercial licence.
License	G0.5 Community License (non-commercial, limited patent licence)	B Hugging Face: "All content committed on or after 2026-06-16 is licensed under the G0.5 Community License (Non-Commercial + Limited Patent License)." Older commits sit under Apache-2.0 or the G0 Plus licence.
The developer’s own claims
Learns new tasks	without examples	C Developer claim: the abstract says "prompts directly steer action granularity, task horizon, and out-of-distribution scene handling without further training". For "two new long-horizon zero-shot tasks": "no parameters are fine-tuned or adapted, and the same weights are reused across all cells". "At runtime the policy receives a per-stage natural-language instruction for the current sub-goal." Galaxea also fine-tunes the model on its robots.
Works across robots	yes	A Developer claim: the base checkpoint is for "Fine-tuning and R1 Lite/R1 Pro zero-shot deployment". "SO-100/101 deployment uses the same G0.5 policy server with a lightweight LeRobot-based client." Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

Galaxea describes G0.5 as a pretrained autoregressive vision-language-action model for general-purpose robot control.
Galaxea introduced the model on 1 June 2026.

Sources

A Manufacturer product page. github.com/OpenGalaxea/GalaxeaVLA · Retrieved 2026-10-14
B huggingface.co/OpenGalaxea/G05 · Retrieved 2026-10-14
C arxiv.org/abs/2608.11739 · Retrieved 2026-10-14

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.