![[2602.21684] Primary-Fine Decoupling for Action Generation in Robotic Imitation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21684] Primary-Fine Decoupling for Action Generation in Robotic Imitation
Summary
The paper presents a two-stage framework, Primary-Fine Decoupling for Action Generation (PF-DAG), aimed at improving action generation in robotic imitation by addressing the limitations of existing methods.
Why It Matters
As robotics increasingly integrates into various industries, enhancing the fidelity and stability of robotic actions is crucial. PF-DAG's innovative approach could lead to more effective and adaptable robotic systems, impacting fields such as automation, manufacturing, and service robotics.
Key Takeaways
- PF-DAG decouples coarse action consistency from fine-grained variations, improving action generation.
- The framework outperforms existing methods across multiple benchmarks, demonstrating its effectiveness.
- Theoretical proofs indicate PF-DAG achieves lower mean squared error than single-stage generative policies.
Computer Science > Robotics arXiv:2602.21684 (cs) [Submitted on 25 Feb 2026] Title:Primary-Fine Decoupling for Action Generation in Robotic Imitation Authors:Xiaohan Lei, Min Wang, Wengang Zhou, Xingyu Lu, Houqiang Li View a PDF of the paper titled Primary-Fine Decoupling for Action Generation in Robotic Imitation, by Xiaohan Lei and 4 other authors View PDF HTML (experimental) Abstract:Multi-modal distribution in robotic manipulation action sequences poses critical challenges for imitation learning. To this end, existing approaches often model the action space as either a discrete set of tokens or a continuous, latent-variable distribution. However, both approaches present trade-offs: some methods discretize actions into tokens and therefore lose fine-grained action variations, while others generate continuous actions in a single stage tend to produce unstable mode transitions. To address these limitations, we propose Primary-Fine Decoupling for Action Generation (PF-DAG), a two-stage framework that decouples coarse action consistency from fine-grained variations. First, we compress action chunks into a small set of discrete modes, enabling a lightweight policy to select consistent coarse modes and avoid mode bouncing. Second, a mode conditioned MeanFlow policy is learned to generate high-fidelity continuous actions. Theoretically, we prove PF-DAG's two-stage design achieves a strictly lower MSE bound than single-stage generative policies. Empirically, PF-DAG outperforms ...