VLA GalleryAll modelsGallery

VLA Gallery / All models / New in 2026

G0.5

Galaxea AI (OpenGalaxea) · 2026-06

VLANew in 2026Total amount not discloseddiscrete tokens

Decoder-only VLA that produces reasoning and actions in a single autoregressive stream.

View architecture diagram in the galleryPaper ↗GitHub ↗

Paper

G0.5: One Autoregressive Stream for Robot Reasoning and ActionarXiv 2608.11739

Overview

Category
New in 2026
Type
VLA
Released
2026-06
Organization
Galaxea AI (OpenGalaxea)
Open weights
Yes
Parameters
Total amount not disclosed (initialized from Qwen3.5 2B VLM)

Architecture

Backbone LLM
Not disclosed
Vision encoder
—
Action head
Autoregressive reasoning·action tokens (learned tokenizer) (discrete tokens)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
A single transformer decoder autoregressively generates reasoning·action tokens under one objective (no separate flow matching expert); learned cross-embodiment action tokenizer, CoT reasoning stream, visual memory module

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
GPU type not disclosed; fine-tuning sentences '4 epochs, LR 4×10⁻⁵, global batch 1024' · '100K steps, LR 1×10⁻⁵' (stage attribution needs checking against the original)
Data type
Real-robot teleoperation (14 embodiments) + web vision-language (web·embodied VQA, VQA:action 1:4)
Data scale
Robot demonstrations from 14 embodiments (total Not disclosed); Galaxea Open-World dataset 500+ hours
Main datasets
Galaxea Open-World Dataset, web · embodied VQA
Tasks & evaluation
DROID zero-shot 82.5%; SimplerEnv-Bridge 87.3%; RoboTwin 2.0 93.3%; LIBERO 98.9%; BEHAVIOR-1K 0.3136; real-robot R1-Lite/R1-Pro 76.7% (π0.5 53.3%, GR00T-N1.7 24.4%)
Robot
14 embodiments incl. Galaxea R1-Lite, R1-Pro

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Robot demonstrations + VQA supervised learning
Base control
Wheeled (Direct base velocity/pose output) · In BEHAVIOR-1K (sim R1 Pro), outputs navigation as a token group separate from manipulation; the real-robot R1 client publishes only arms · torso · gripper

Notes

Key features
Decoder-only VLA that produces reasoning and actions in a single autoregressive stream. Scored 76.7% on real R1-Lite/R1-Pro robots, ahead of π0.5 (53.3%) and GR00T-N1.7 (24.4%).
Notes
Announced 2026-06-01, base checkpoint 2026-06-16, arXiv 2026-08-12. G0.5 Community License (non-commercial, commits from 2026-06-16 onward; three-tier license). The note's 'about 2B' only reflects 'initialized from Qwen3.5 2B' in the source; there is no total-size sentence. Galaxea is a different company from Galbot.
Verification
Confirmed

Other models in this category