收录解读
现有的视觉-语言-动作(VLA)模型通常采用预训练VLM加独立流匹配动作专家的组合,导致VLM仅作为上下文编码器而非决策者。该论文提出G0.5,一个预训练的自回归VLA模型,用单一Transformer解码器在统一目标下生成推理和动作token,旨在解决VLA模型决策能力不足的问题。
G0.5的核心方法包括三个组件:可学习的跨具身动作tokenizer,将异构机器人动作映射为共享词汇;原生思维链流,将任务分解、物体接地和动作提示与动作token交错;以及视觉记忆模块,通过视觉编码器注入多秒历史信息。由于推理和动作共享同一组权重,预训练VLM的能力直接迁移到物理行为,且无需额外训练即可通过提示控制动作粒度、任务周期和分布外场景处理。
模型在大量机器人数据集和VQA样本上预训练,在七个独立基准上超越当前最优:真实世界微调(76.7% vs 53.3% π0.5 和 24.4% GR00T-N1.7)、BEHAVIOR挑战(31.4% vs 26.3% π0.5 和 26.1% 冠军)、DROID后训练零样本迁移(82.5%)、LIBERO(98.9%)、RoboTwin 2.0(93.3%)、SimplerEnv-Bridge(87.3%)等。实验覆盖长时程操作、跨具身泛化和语言跟随任务。
G0.5展示了统一自回归架构在机器人推理和动作中的强大潜力,其跨具身tokenizer和思维链设计为VLA模型提供了新的可复用范式。但论文为预印本,未经同行评审,且方法原创性尚不足以称为范式级突破;未来工作需进一步验证其可扩展性和在更广泛真实场景中的鲁棒性。
原始摘要
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 $π_{0.5}$ 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 $π_{0.5}$ 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\%).