G0.5: Single Autoregressive Stream for Robot Reasoning and Action
A recent publication on arXiv (2608.11739) presents G0.5, an autoregressive Vision-Language-Action (VLA) model that integrates reasoning and action generation within a unified transformer decoder. In contrast to traditional VLA models that combine a pretrained Vision-Language Model (VLM) with a separately trained action expert, G0.5 positions the VLM as a decision-maker instead of merely a context encoder. The model generates both reasoning and action tokens from a single objective, facilitated by three innovative components: a learnable action tokenizer that aligns various robot actions with a common vocabulary, a native chain-of-thought stream that merges task decomposition, object grounding, and action hints with action tokens, and a visual memory module that incorporates multi-second history via the vision encoder. This shared weight structure allows the pretrained VLM's skills to enhance physical actions, enabling the model to adhere closely to instructions and adjust action granularity through prompts. The paper is categorized as a cross-type submission on arXiv.
Key facts
- G0.5 is a pretrained autoregressive VLA model.
- It uses a single transformer decoder for reasoning and action tokens.
- It departs from the prevailing recipe of coupling a VLM with a separate action expert.
- Three components: cross-embodiment action tokenizer, chain-of-thought stream, visual memory module.
- The action tokenizer maps heterogeneous robot actions into a shared vocabulary.
- The chain-of-thought stream interleaves task decomposition, object grounding, and action hints.
- Visual memory injects multi-second history through the vision encoder.
- Reasoning and action share weights, enabling VLM capabilities to carry over to physical behavior.
Entities
Institutions
- arXiv