VLA Gallery / All models / Industry models

Galaxea G0

Galaxea · 2025-08

VLAIndustry modelsNot disclosedNot disclosed

A two-tier design where a VLM planner handles high-level instructions and a VLA executor carries them out.

Paper

Galaxea Open-World Dataset and G0 Dual-System VLA ModelarXiv 2509.00576

Overview

Category

Industry models

Type

VLA

Released

2025-08

Organization

Galaxea

Open weights

Yes

Parameters

Not disclosed

Architecture

Backbone LLM

Gemma (PaliGemma, execution VLA) + Qwen2.5-VL (planner)

Vision encoder

SigLIP + MLP projector

Action head

Not disclosed (Not disclosed)

Chunk

Not disclosed

Control rate

200 Hz (planning 15 Hz)

Architecture (original summary)

dual-system: G0-VLM (Qwen2.5-VL-based planner, System 2) + G0-VLA (SigLIP + MLP projector + PaliGemma language backbone executor, System 1); control 200Hz·planning 15Hz

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Not disclosed

Data type

Real robot + public robot data (OXE)

Data scale

Galaxea Open-World Dataset 500 h, 100K traj, 150 task categories, 50 scenes, 58 skills; training mix OXE about 1,000 h + Galaxea 500 h + in-house 200 h

Main datasets

Galaxea Open-World Dataset, Open X-Embodiment, in-house data

Tasks & evaluation

Table Bussing, Microwave Operation, Bed Making (4-point scale), Blocks Stacking (language following); quantitative figures not extracted

Robot

Galaxea R1 Lite (23-DoF, wheeled · torso bimanual mobile)

Inference

Inference device & speed

Control 200Hz, planning 15Hz (per figure); hardware and latency not disclosed

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not used · Imitation learning (behavior cloning) only

Base control

Wheeled (Direct base velocity/pose output) · G0-VLA action chunks control chassis · torso · arms together (chassis moves during bed making); chassis command format not stated

Notes

Key features

A two-tier design where a VLM planner handles high-level instructions and a VLA executor carries them out. Uses a 3-stage curriculum: cross-embodiment pretraining → single-embodiment pretraining → post-training.

Notes

arXiv v1 2025-08-30 (corrected from 2025-09 in notes). G0-VLA weights released 2025-09-09 on HF OpenGalaxea/G0-VLA·ModelScope (corrected from Unverified in notes; paper says planned release under CC BY-NC-SA 4.0). Data hours (h) based on Galaxea Open-World Dataset (no statement of training mixture total). Paper has no statements on GPU·parameters.

Verification

Corrections applied