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

2026-08-12

AI Summary

This work addresses the limitation of existing vision-language-action (VLA) models, which treat vision-language models (VLMs) merely as contextual encoders, thereby decoupling reasoning from action generation. The authors propose G0.5, the first end-to-end pretrained autoregressive VLA model, which unifies reasoning and action generation within a single Transformer decoder for joint optimization. Key innovations include a learnable cross-embodied action tokenizer, a native chain-of-thought stream that interleaves task decomposition, object grounding, and action prompting, and a visual memory module supporting multi-second history. This architecture enables direct transfer of VLM capabilities to physical behavior control. G0.5 achieves state-of-the-art performance across seven benchmarks, including real-world robots R1lite/R1pro (76.7%), BEHAVIOR Challenge (31.4%), zero-shot DROID transfer (82.5%), and LIBERO (98.9%).

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 $π_{0.5}$ 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\%).

Problem

• Vision-Language-Action
• robot reasoning
• autoregressive modeling
• action generation
• foundation models

Innovation

• autoregressive VLA
• cross-embodiment action tokenizer
• chain-of-thought reasoning
• visual memory module
• foundation robot model