FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency

TL;DR

FreqPolicy addresses the latency of flow-based visuomotor policies by enforcing temporal structure through frequency-domain constraints, enabling high-quality one-step action generation for robotic manipulation. It introduces a frequency consistency objective to align velocity signals across timesteps and an adaptive frequency component loss to capture structured temporal variations, integrated with a flow-matching backbone. Empirical results across 53 simulated tasks and real-world LIBERO/VLA evaluations show improved one-step performance and substantial speedups (e.g., up to 5x faster inference) with minimal or no loss in task success, including real-world demonstrations at 93.5 Hz. The approach significantly narrows the gap between one-step generators and multi-step policies, with practical impact for real-time embodied AI systems and vision-language-action frameworks.

Abstract

Generative modeling-based visuomotor policies have been widely adopted in robotic manipulation, attributed to their ability to model multimodal action distributions. However, the high inference cost of multi-step sampling limits its applicability in real-time robotic systems. Existing approaches accelerate sampling in generative modeling-based visuomotor policies by adapting techniques originally developed to speed up image generation. However, a major distinction exists: image generation typically produces independent samples without temporal dependencies, while robotic manipulation requires generating action trajectories with continuity and temporal coherence. To this end, we propose FreqPolicy, a novel approach that first imposes frequency consistency constraints on flow-based visuomotor policies. Our work enables the action model to capture temporal structure effectively while supporting efficient, high-quality one-step action generation. Concretely, we introduce a frequency consistency constraint objective that enforces alignment of frequency-domain action features across different timesteps along the flow, thereby promoting convergence of one-step action generation toward the target distribution. In addition, we design an adaptive consistency loss to capture structural temporal variations inherent in robotic manipulation tasks. We assess FreqPolicy on 53 tasks across 3 simulation benchmarks, proving its superiority over existing one-step action generators. We further integrate FreqPolicy into the vision-language-action (VLA) model and achieve acceleration without performance degradation on 40 tasks of LIBERO. Besides, we show efficiency and effectiveness in real-world robotic scenarios with an inference frequency of 93.5 Hz.

Figures (5)

• Figure 1: Overview of FreqPolicy. a) For training, we use the frequency consistency constraint to align the velocity vectors across different time steps in the frequency space. Besides, we introduce an adaptive frequency component loss to accommodate the diverse frequency structures in manipulation tasks. b) FreqPolicy takes 2D or 3D input and predicts the velocity vector of the action as output.
• Figure 2: Experimental benchmarks. We evaluate FreqPolicy on 5 benchmarks, including a total of $93$ simulated tasks (left) and $3$ real robotics tasks (right).
• Figure 3: Demonstrations of three real-world tasks (top) and the test results of different policies (bottom). We evaluated the success rate (%) and inference speed (Hz) for all methods. DP stands for Diffusion Policy, FM for Flow Matching Policy, and the number represents inference steps of the corresponding policy model. FreqPolicy consistently outperformed baselines in both success rate and inference speed, demonstrating its effectiveness on real-world robotic platforms.
• Figure 4: Visualization of the spectral and temporal signals across different action chunks.
• Figure 5: Demonstrations of three long-horizon real-world tasks. Following a consistent protocol, 300 episodes are collected for each task for training, after which the trained policies are evaluated. The numbers in each image indicate the sequential steps in the task execution process.