← 首页|学术|G0.5: One Autoregressive Stream for Robot Reasoning and Action
cs.RO / cs.AI · 2608.11739 · 2026/08/12

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

Liu, Yicheng, Dong, Zibin, Ye, Baijun, Yuan, Tianyuan, Jiang, Tao, Yang, Anqi, Cao, Shicheng, Liu, Haonan
📄 摘要🌐 HTML📑 PDF
TL;DR：主流VLA是「预训练VLM + 单独flow-matching动作专家」，VLM沦为上下文编码器而非决策者。G0.5用单一transformer decoder在单一目标下同时发射推理token与动作token，推理与动作共享一套权重，VLM预训练能力直接传导到物理行为。

🎯 问题

VLA主流配方让VLM成为上下文编码器，推理与动作分离，决策能力受限。

🔬 方法

可学习跨具身动作tokenizer（异构机器人动作映射到共享词表）+ 原生CoT流（任务分解/物体grounding/动作提示与动作token交错）+ 视觉记忆模块（多秒历史经视觉编码器注入）；单一目标自回归。
章节结构（全文标题提取）：
1 Introduction
2 Related Work
3 G0.5 Model Design
4 G0.5 Pre-training
5 Experiments
6 Conclusion
7 Contributors
References
2.1 VLA Architectures: from VLM-as-Encoder to VLM-as-Actor
2.2 Action Tokenization and Cross-Embodiment
2.3 Reasoning and Chain-of-Thought in VLAs
3.1 Structured Tokenization of Heterogeneous Action Data
3.2 Native Chain-of-Thought
3.3 Visual Memory
5.1 DROID Environment- and Object-Level Zero-Shot Evaluation
5.2 Simulation Benchmarks
5.3 Long-horizon Tasks
5.4 Real-World Fine-Tuning Evaluation
5.5 Pick-and-Place Benchmark
5.6 Zero-Shot Probe of CoT and Action Head

📊 结果

R1lite/R1pro真机微调76.7%（vs π0.5的53.3%、GR00T-N1.7的24.4%）；BEHAVIOR Challenge 2025（50个长时程家务任务）31.4% vs π0.5的26.3%；DROID后训练零样本迁移82.5%。

💡 与研究方向关联

推理与动作共享一套权重的单流 VLA——VLM 从上下文编码器变回决策者。对具身 agent 架构的统一化趋势有代表性意义。

📝 原文摘要

▶ 原文摘要 Abstract
The prevailing recipe for Vision-Language-Action (VLA) models couples a pretrained VLM with a separately trained flow-matching action expert. This makes the VLM a context encoder rather than a decision-maker. We introduce G0.5, a pretrained autoregressive VLA in which a single transformer decoder emits reasoning and action tokens under a single objective. Three components make this tractable at foundation-model scale: a learnable cross-embodiment action tokenizer that maps heterogeneous robot actions into a shared vocabulary; a native chain-of-thought stream interleaving task decomposition, object grounding, and action hints with action tokens; and a visual memory module that injects multi-second history through the vision encoder. Because reasoning and action share a single set of weights, the pretrained VLM's capabilities carry over to physical behavior: the model follows instructions closely, and prompts directly steer action granularity, task horizon, and out-of-distribution scene handling without further training. Pretrained on a large collection of robot datasets together with VQA samples, G0.5 surpasses state-of-the-art models across 7 independent regimes: real-world fine-tuning on R1lite and R1pro robots (76.7\% vs.\ 53.3\% for $\pi_{0.5}$ and 24.4\% for GR00T-N1.7), the 2025 BEHAVIOR Challenge on 50 long-horizon household mobile manipulation tasks using a generalist policy (31.4\% vs.\ 26.3\% for $\pi_{0.5}$ and 26.1\% for the challenge winner), DROID post-training followed by zero-shot transfer to an unseen environment and objects (82.5\%), a language-following Pick-and-Place benchmark, LIBERO (98.9\%), RoboTwin 2.0 (93.3\%), and SimplerEnv-Bridge (87.3\%).
Deep Read · 2026-08-14高松灯 / Agent 日报
VLAAutoregressiveEmbodied AIAction Tokenizer