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

1. Problem Statement

Current approaches to general-purpose robotics often rely on dual-system frameworks inspired by Kahneman's cognitive theory:

• System Two (Reasoning): A high-level Vision-Language Model (VLM) performs slow, complex reasoning and planning.
• System One (Acting): A Vision-Language-Action (VLA) model executes low-level robot actions based on the reasoning output.

Limitations of Dual-System Approaches:

• Lack of Mutual Awareness: The reasoning system may generate plans that the acting system cannot physically execute due to capability mismatches.
• Latency: The decoupled nature causes delays; the reasoning system may respond too late, providing outdated guidance for dynamic environments.
• Inefficiency: "Always reasoning" approaches incur significant inference overhead, while "no reasoning" approaches lack robustness in long-horizon tasks.

The paper argues that achieving true synergistic reasoning and acting requires a single unified model that can adaptively switch between reasoning and acting modes.

2. Methodology: OneTwoVLA

Core Architecture:
OneTwoVLA is a unified VLA model capable of both System One (Acting) and System Two (Reasoning) behaviors within a single architecture. It is instantiated based on $\pi_0$ (a state-of-the-art flow-matching VLA).

Adaptive Inference Pipeline:
The model autonomously decides at each timestep whether to reason or act using two special decision tokens:

1. [BOR] (Beginning of Reasoning): Triggers the model to generate textual reasoning content (Scene description, Plan, Historical Summary, Next Step).
2. [BOA] (Beginning of Action): Triggers the model to generate an action chunk ($A_t$) based on the most recent reasoning.

• Mechanism: The model takes current observations ($I_t$), reference images from the last reasoning step ($I_{ref}$), the language instruction ($\ell$), and the latest reasoning content ($R$). It predicts either [BOR] or [BOA].
• Efficiency: The model operates primarily in acting mode, invoking reasoning only at critical moments (e.g., subtask completion, error detection, or ambiguity), minimizing latency.

Data Curation & Synthesis:
To train the model for this dual capability, the authors developed two data strategies:

1. Embodied Reasoning Robot Data:
   • Human demonstration trajectories are segmented into Reasoning Intervals (key steps requiring updates) and Acting Intervals (continuous execution).
   • Reasoning content includes: Scene description, High-level plan, Historical summary, and Next-step instruction.
   • Includes specific data for error recovery and human-robot interaction (HRI).
2. Scalable Synthetic Vision-Language Data:
   • A pipeline using Gemini 2.5 Pro and FLUX.1-dev generates 16,000 synthetic samples.
   • Process: Text-to-image generation of tabletop layouts $\rightarrow$ Gemini generates task instructions and reasoning content.
   • Types: Visual grounding tasks (spatial/semantic/attribute) and Long-horizon planning tasks.
   • Augmentation: Images are augmented with fisheye distortion and robot gripper compositing to mimic real robot observations.

Training Strategy:

• Co-training: The model is trained jointly on the curated robot data and the synthetic vision-language data.
• Loss Functions: Cross-entropy for text/reasoning generation and Flow Matching loss for continuous action prediction.
• Input Context: Includes current images, reference images (to prevent confusion from outdated text descriptions), and temporal proprioceptive states.

3. Key Contributions

1. Unified Adaptive Framework: Introduction of OneTwoVLA, the first unified VLA that dynamically switches between reasoning and acting modes, eliminating the latency and capability gaps of dual-system approaches.
2. Embodied Reasoning Data Pipeline: A novel method for curating robot data with explicit reasoning intervals and a scalable pipeline for synthesizing high-quality, embodied reasoning-centric vision-language data.
3. Superior Performance: Demonstrated state-of-the-art results in four critical capabilities:
   • Long-horizon task planning.
   • Real-time error detection and recovery.
   • Natural human-robot interaction (HRI).
   • Generalizable visual grounding (including unseen objects).

4. Experimental Results

The model was evaluated on a Franka arm and a dual-arm ARX platform across various tasks (e.g., making hotpot, mixing cocktails, cooking tomato-egg).

• Long-Horizon Planning:
  • OneTwoVLA achieved an 87% success rate, outperforming the baseline VLA ($\pi_0$) by 30% and the Dual-System approach by 24%.
  • It successfully tracked task progress and adjusted plans dynamically, whereas baselines often got stuck or lost context.
• Error Detection & Recovery:
  • Achieved 80% success in recovering from errors (e.g., failed grasps, slipping bottles), compared to 57.1% for $\pi_0$ and 58.3% for the Dual-System.
  • The unified model could reason about the error and execute a corrective strategy immediately, while the Dual-System was too slow to recover before the robot entered an out-of-distribution state.
• Human-Robot Interaction (HRI):
  • Achieved 100% success in handling interruptions and clarifying ambiguities (e.g., "I want lemon vodka instead of orange").
  • The Dual-System struggled to maintain context during interaction, often failing to update the plan correctly.
• Generalizable Visual Grounding:
  • In Open-World settings (unseen objects/environments), OneTwoVLA-VL (co-trained with synthetic data) achieved 73% success, significantly outperforming $\pi_0$ (3%) and OneTwoVLA without synthetic data (8%).
  • It successfully generalized to objects never seen in robot training data (e.g., Sprite, GoPro) by leveraging knowledge from the synthetic vision-language data.
• Efficiency:
  • Inference time in acting mode is comparable to $\pi_0$ (~0.1s).
  • Reasoning occurs only ~5 times per long-horizon task, adding minimal overhead (e.g., 16s of reasoning in a 183s task).

5. Significance and Future Work

Significance:
OneTwoVLA demonstrates that unifying reasoning and acting in a single model, supported by adaptive inference and co-training with synthetic embodied data, is a superior paradigm for general-purpose robotics. It bridges the gap between high-level cognitive planning and low-level motor control, enabling robots to handle complex, dynamic, and interactive real-world tasks with human-like flexibility.

Limitations & Future Directions:

• Heuristic Triggers: Current reasoning triggers are hand-selected; future work could use Reinforcement Learning (RL) to optimize when to reason.
• Latency: The robot must pause during reasoning; asynchronous architectures could allow simultaneous reasoning and acting.
• Scalability: As model size increases, action inference efficiency may become a bottleneck, requiring advanced LLM inference techniques.
• Data Sources: Future work could explore diverse sources of vision-language data beyond the current synthetic pipeline.