Robotics

G0.5: One Autoregressive Stream for Robot Reasoning and Action

Yicheng Liu, Zibin Dong

Featured August 23, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

This paper introduces a new robot brain, G0.5, that thinks and acts all at once like a storyteller, using a single stream of "words" to plan steps and control different robots, making it much better at following instructions.

In depth

The paper introduces G0.5, a unified Vision-Language-Action (VLA) model that integrates reasoning and action generation into a single autoregressive token stream. Unlike prior approaches that separate the VLM from an action expert, G0.5 uses a single transformer decoder with a shared objective, enabling native chain-of-thought reasoning and direct prompt-driven control over physical robot behavior.

Key Takeaways

1
Unified Autoregressive Stream: G0.5 unifies perception, reasoning, and action within a single autoregressive process, allowing the VLM to directly generate both intermediate thoughts and robot commands.
2
Cross-Embodiment Action Codec: The model employs a learnable action tokenizer that maps heterogeneous robot actions into a shared, structured token vocabulary, making autoregressive VLA practical across diverse robot morphologies.
3
Native Chain-of-Thought: G0.5 integrates chain-of-thought reasoning (subtask decomposition, object grounding, action hints) directly into the generative token stream, improving instruction following and action accuracy without separate modules.

Conceptual Flow

HIGH LEVEL

Methodology (The "Logic")

Instead of two separate brains for thinking and acting, this robot uses one smart brain that does both at the same time.

Robot Sees

Robot Hears

One Smart Brain

Robot Thinks

Robot Acts

Results (The "Impact")

This new brain helps robots understand instructions better and perform complex tasks more reliably than older methods.

Old Robot Brain

New Robot Brain

Better Understanding

More Success

Harder Tasks

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.