OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning

Fanqi Lin, Ruiqian Nai, Yingdong Hu, Jiacheng You, Junming Zhao, Yang Gao

Abstract

General-purpose robots capable of performing diverse tasks require synergistic reasoning and acting capabilities. However, recent dual-system approaches, which separate high-level reasoning from low-level acting, often suffer from challenges such as limited mutual understanding of capabilities between systems and latency issues. This paper introduces OneTwoVLA, a single unified vision-language-action model that can perform both acting (System One) and reasoning (System Two). Crucially, OneTwoVLA adaptively switches between two modes: explicitly reasoning at critical moments during task execution, and generating actions based on the most recent reasoning at other times. To further unlock OneTwoVLA's reasoning and generalization capabilities, we design a scalable pipeline for synthesizing embodied reasoning-centric vision-language data, used for co-training with robot data. We validate OneTwoVLA's effectiveness through extensive experiments, highlighting its superior performance across four key capabilities: long-horizon task planning, error detection and recovery, natural human-robot interaction, and generalizable visual grounding, enabling the model to perform long-horizon, highly dexterous manipulation tasks such as making hotpot or mixing cocktails.

Introduction

In robotics, the integration of high-level cognitive reasoning (System Two) with low-level motor control (System One) remains a significant hurdle. Traditional architectures typically decouple these systems, employing a Large Language Model (LLM) for planning and a separate Vision-Language-Action (VLA) model for execution. This separation often leads to a "communication gap," where the high-level planner lacks physical grounding, and high latency, which prevents effective real-time error recovery.

OneTwoVLA addresses these challenges by unifying reasoning and acting within a single transformer-based model. By adaptively switching between modes, the model can "pause and think" at critical junctures while maintaining high-frequency control during routine execution.

OneTwoVLA adaptively interleaves reasoning and acting modes during task execution, allowing it to update its plan based on the current scene and history.

Methodology

The core innovation of OneTwoVLA is its Adaptive Reasoning framework. Unlike "flat" VLAs that map images directly to actions, OneTwoVLA utilizes specialized tokens to manage its operational state:

  • [BOR] (Beginning of Reasoning): Triggers a reasoning mode where the model generates a structured text block containing a scene description, a high-level plan, a historical summary, and the immediate next step.
  • [BOA] (Beginning of Action): Triggers the acting mode, where the model outputs raw motor commands (action chunks) based on the most recent reasoning.

To train this system, the researchers developed an automated labeling pipeline using Gemini 2.5 Pro to segment human demonstrations into reasoning and acting intervals. Furthermore, they utilized a scalable pipeline to generate 16,000 synthetic reasoning-centric samples, leveraging FLUX.1-dev for high-quality imagery to enhance visual grounding.

Experimental Results

OneTwoVLA was evaluated across several long-horizon tasks, including cooking and cocktail mixing, as well as open-world generalization benchmarks involving 180 distinct household items.

Performance and Error Recovery

The model achieved an 87% average success rate on complex tasks, outperforming standard VLAs by 30% and dual-system models by 24%. One of its most significant strengths is error recovery; OneTwoVLA successfully corrected mistakes, such as dropped objects or missed grasps, in 80% of cases by re-evaluating its plan when visual feedback indicated a failure.

Efficiency and Latency

A primary drawback of dual-system approaches is the latency penalty incurred by constant reasoning. OneTwoVLA mitigates this by reasoning only at "critical moments," such as when a subtask is completed or an error is detected.

Comparison of task completion times. OneTwoVLA achieves reasoning capabilities with significantly lower latency than dual-system approaches.

As shown in the comparison of task times, OneTwoVLA (184s) is significantly faster than dual-system architectures (260s) and nearly as efficient as "mindless" models (176s) while providing superior task success and reasoning capabilities.

Conclusion

OneTwoVLA represents a major step toward robots that possess both the intelligence to plan and the dexterity to act. By unifying these capabilities and employing adaptive reasoning, the model overcomes the latency and communication barriers of previous systems. Future research will explore Reinforcement Learning to further optimize the model's "thinking schedule" and enhance its precision in delicate manipulation tasks.

Summary updated

Abstract

General-purpose robots capable of performing diverse tasks require synergistic reasoning and acting capabilities. However, recent dual-system approaches, which separate high-level reasoning from low-level acting, often suffer from challenges such as limited mutual understanding of capabilities between systems and latency issues. This paper introduces OneTwoVLA, a single unified vision-language-action model that can perform both acting (System One) and reasoning (System Two). Crucially, OneTwoVLA adaptively switches between two modes: explicitly reasoning at critical moments during task execution, and generating actions based on the most recent reasoning at other times. To further unlock OneTwoVLA's reasoning and generalization capabilities, we design a scalable pipeline for synthesizing embodied reasoning-centric vision-language data, used for co-training with robot data. We validate OneTwoVLA's effectiveness through extensive experiments, highlighting its superior performance across four key capabilities: long-horizon task planning, error detection and recovery, natural human-robot interaction, and generalizable visual grounding, enabling the model to perform long-horizon, highly dexterous manipulation tasks such as making hotpot or mixing cocktails.
BibTeX
@inproceedings{lin2026onetwovla,
  author = {Fanqi Lin and Ruiqian Nai and Yingdong Hu and Jiacheng You and Junming Zhao and Yang Gao},
  title = {{OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=tWMfhoP3as},
}
Figure