Galaxea G0 Dual-System VLA

Galaxea's 2025 open-source fast-slow dual-system robot model, where a VLM breaks down the task and a VLA executes the actions.

G0 is a dual-system robot foundation model from the embodied-AI company Galaxea (Galaxea AI), released in late August 2025 and open-sourced in September, launched alongside the Galaxea Open-World Dataset. It splits into a fast-slow dual system: the slow system, G0-VLM, is fine-tuned from Qwen2.5-VL and handles understanding instructions and breaking a long task into sub-tasks; the fast system, G0-VLA, has a vision-language backbone initialized from PaliGemma and outputs actions based on the current view, the robot's own state, and the sub-task instruction. Training has three stages: cross-embodiment pretraining on data from multiple robots including OXE, then pretraining on the company's own single-embodiment data, and finally post-training on specific tasks. The paper found the second stage — pretraining on real data from the same type of robot — to be the most important. The dataset was collected with the Galaxea R1 Lite across 50 real-world scenes spanning homes, retail, dining, and offices, totaling about 500 hours and 100,000 trajectories. Galaxea later released G0Plus (January 2026) and G0.5 (June 2026).

ExampleWhen the user says “help me make the bed,” G0-VLM breaks it into a sequence of sub-task instructions, and G0-VLA drives the R1 Lite's two arms and mobile base step by step according to those instructions, until the bed is made.

Also called

G0, G0-VLA, Galaxea G0

Related

Galaxea AI · Dual-System Architecture (System 1 / System 2) · Galaxea Open-World Dataset · Galaxea R1 · PaliGemma · Vision-Language-Action Model

Sources

Galaxea Open-World Dataset and G0 Dual-System VLA Model (arXiv 2509.00576)
OpenGalaxea/GalaxeaVLA（GitHub）

As of

2026-08