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

2025-05-17

AI Summary

Existing general-purpose robots commonly adopt a decoupled “dual-system” architecture—separating action execution (System I) from symbolic reasoning (System II)—leading to poor cross-system capability alignment and response latency. This paper introduces the first end-to-end unified vision-language-action model, integrating perception, reasoning, and action within a single Transformer framework. A dynamic reasoning gating mechanism adaptively triggers explicit, step-wise reasoning only at task-critical junctures; otherwise, actions are generated directly from prior reasoning outcomes. Key innovations include: (1) a dual-modal adaptive switching mechanism for seamless perception–reasoning coordination; (2) a scalable synthetic triplet data generation pipeline tailored for embodied reasoning; and (3) a multi-stage joint fine-tuning strategy. Experiments demonstrate substantial improvements over baselines across four core capabilities—long-horizon planning, error recovery, natural human–robot interaction, and generalizable spatial localization—enabling successful execution of highly dexterous, multi-step tasks such as hotpot preparation and cocktail mixing.

Technology Category

• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Computer Vision: Visual Reasoning & Symbolic Representations
• Natural Language Processing: Safety and Robustness

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Responsible Web: Machine-in-the-loop, human agency and autonomy

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.

Problem

• Unified model for vision-language-action with adaptive reasoning
• Overcoming dual-system limitations in robot task execution
• Enhancing reasoning and generalization for diverse robotic tasks

Innovation

• Unified vision-language-action model with adaptive reasoning
• Scalable pipeline for synthesizing reasoning-centric data
• Co-training with robot data for enhanced generalization