Dense-Jump Flow Matching with Non-Uniform Time Scheduling for Robotic Policies: Mitigating Multi-Step Inference Degradation

Zidong Chen, Zihao Guo, Peng Wang, ThankGod Itua Egbe, Yan Lyu, Chenghao QianarXiv 2025 · DOI · arXiv

Introduction

Generative policies, particularly those based on Diffusion models, have significantly advanced robot learning by effectively capturing complex, multi-modal action distributions. However, Diffusion models are computationally expensive due to their iterative sampling process. Flow Matching (FM) has emerged as a promising, faster alternative that learns a deterministic velocity field to transform noise into expert actions.

Despite its success in image generation, this paper identifies a critical and counter-intuitive failure mode for FM in robotics: multi-step degradation. Unlike image generation, where more inference steps typically lead to higher quality, FM robotic policies often peak at 2–4 steps and then suffer a sharp decline in performance as the number of steps increases.

The Multi-Step Degradation Problem

The authors identify two primary causes for this performance collapse:

  1. Mathematical Instability (Lipschitz Blow-up): The paper proves that the Lipschitz constant of the ideal FM velocity field scales as $1/(1-t)$. As the integration time $t$ approaches the target ($t \to 1$), the system becomes hyper-sensitive to infinitesimal errors. This causes the ODE integration to become unstable, leading to "blown-up" trajectories.
  2. Localised Overfitting: Standard FM training uses uniform time sampling, which over-indexes on the middle of the trajectory. In continuous robotic states that never perfectly repeat, this causes the model to "drift" toward specific training actions (memorization) rather than generalizing to the current observation (conditional faithfulness).
Reward vs. Inference Steps: Vanilla FM performance plummets as steps increase, while FM-DJβ remains robust.

Proposed Method: FM-DJβ

To address these issues, the authors introduce Dense-Jump Flow Matching with Beta-scheduling (FM-DJβ), a framework consisting of two synergistic components:

Non-Uniform Time Scheduling (Training)

Instead of uniform sampling, the authors propose using a Beta distribution to create a U-shaped time schedule during training. By oversampling $t \approx 0$, the model improves its "conditional faithfulness," anchoring the action more strongly to the robot's observation. Oversampling $t \approx 1$ acts as a regularizer, forcing the neural network to learn a smoother, lower-frequency function in the high-curvature region.

Dense-Jump ODE Solver (Inference)

To bypass the instability near the target, the authors propose a Dense-Jump strategy. Instead of taking many small steps all the way to $t=1$, the solver takes $N-1$ steps through the stable early region and then performs a single, large "jump" from a predefined switch point (e.g., $t=0.5$) directly to the final action. This avoids the non-Lipschitz region where errors accumulate most rapidly.

Conceptual overview of the Dense-Jump strategy, which bypasses the unstable, high-curvature region near the target action.

Experimental Evaluation

The method was evaluated on three challenging MuJoCo-based benchmarks:

  • Walker2D-v5: A locomotion task requiring stable, continuous forward movement.
  • AdroitHandPenSparse-v1: A complex dexterous manipulation task involving a 24-DOF Shadow Hand.
  • HumanoidStandup-v5: A high-dimensional (348-dim state) task where a humanoid must stand up from a prone position.

Key Results

  • Robustness to Step Count: While vanilla FM saw success rates on Walker2D drop from 84.6% (at 2 steps) to 66.6% (at 64 steps), FM-DJβ maintained a consistent success rate of ~87-89% regardless of the number of steps used.
  • Superior One-Step Performance: FM-DJβ significantly improves single-step inference. On the Adroit Pen task, the reward increased by 16.6% over the baseline, making it highly suitable for real-time applications.
  • State-of-the-Art Gains: The method achieved up to 23.7% performance gains over existing baselines and consistently outperformed Diffusion policies, which often fail entirely at low step counts (1–4 steps).

Conclusion

FM-DJβ effectively mitigates the multi-step degradation inherent in standard Flow Matching for robotics. By combining a U-shaped training schedule with a Dense-Jump inference solver, the method achieves a robust, high-performance generative policy that generalizes better and remains stable across varying computational budgets. Future work will explore adaptive jump points and the impact of real-world hardware noise on the "jump" strategy.

Summary updated

BibTeX
@misc{chen2025dense,
  author = {Zidong Chen and Zihao Guo and Peng Wang and ThankGod Itua Egbe and Yan Lyu and Chenghao Qian},
  title = {{Dense-Jump Flow Matching with Non-Uniform Time Scheduling for Robotic Policies: Mitigating Multi-Step Inference Degradation}},
  year = {2025},
  eprint = {2509.13574},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.13574},
}
Figure