![[2602.13718] HybridFlow: A Two-Step Generative Policy for Robotic Manipulation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.13718] HybridFlow: A Two-Step Generative Policy for Robotic Manipulation
Summary
The paper presents HybridFlow, a two-step generative policy designed to improve robotic manipulation by enhancing real-time interaction capabilities while reducing inference time significantly.
Why It Matters
As robotics increasingly integrates into real-world applications, the ability to perform tasks with low latency and high precision is critical. HybridFlow addresses the limitations of existing methods, offering a promising solution that balances speed and accuracy, which is essential for advancing robotic manipulation technologies.
Key Takeaways
- HybridFlow improves robotic manipulation success rates by 15-25% compared to traditional methods.
- The method achieves an 8x reduction in inference time, enhancing real-time interaction.
- HybridFlow combines global and local refinement strategies for optimal performance.
- The approach shows strong potential for handling unseen and deformable objects.
- This innovation could significantly impact the deployment of robots in dynamic environments.
Computer Science > Robotics arXiv:2602.13718 (cs) [Submitted on 14 Feb 2026] Title:HybridFlow: A Two-Step Generative Policy for Robotic Manipulation Authors:Zhenchen Dong, Jinna Fu, Jiaming Wu, Shengyuan Yu, Fulin Chen, Yide Liu View a PDF of the paper titled HybridFlow: A Two-Step Generative Policy for Robotic Manipulation, by Zhenchen Dong and 5 other authors View PDF HTML (experimental) Abstract:Limited by inference latency, existing robot manipulation policies lack sufficient real-time interaction capability with the environment. Although faster generation methods such as flow matching are gradually replacing diffusion methods, researchers are pursuing even faster generation suitable for interactive robot control. MeanFlow, as a one-step variant of flow matching, has shown strong potential in image generation, but its precision in action generation does not meet the stringent requirements of robotic manipulation. We therefore propose \textbf{HybridFlow}, a \textbf{3-stage method} with \textbf{2-NFE}: Global Jump in MeanFlow mode, ReNoise for distribution alignment, and Local Refine in ReFlow mode. This method balances inference speed and generation quality by leveraging the rapid advantage of MeanFlow one-step generation while ensuring action precision with minimal generation steps. Through real-world experiments, HybridFlow outperforms the 16-step Diffusion Policy by \textbf{15--25\%} in success rate while reducing inference time from 152ms to 19ms (\textbf{8$\times$ ...
