Research map / Image generation / Fast diffusion sampling
FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FreqPolicy, a one-step flow-based visuomotor policy that adds a DCT frequency-domain velocity consistency loss with an adaptive per-frequency weighting, and reports gains on Robomimic, Adroit/MetaWorld, LIBERO (with OpenVLA), and real robots.
Problem
Multi-step sampling in generative visuomotor policies causes high inference latency, and existing one-step accelerators borrowed from image generation ignore the temporal structure of action chunks.
Approach
Treat each action chunk as a temporal signal: enforce consistency between flow velocities at different timesteps in the DCT frequency domain, weighted adaptively per frequency band (Focal-loss-inspired), so one-step generation converges to the target distribution without a teacher model.
Key results
- On 53 Adroit and MetaWorld tasks, FreqPolicy with 1 NFE reaches 78.5% average success, versus 77.2% for FlowPolicy and 74.8% for SDM.
- With OpenVLA on LIBERO (40 tasks), OpenVLA-FreqPolicy reaches 94.8% average success at 1 NFE and 6.05 Hz, versus 93.7% for 10-step flow matching at 1.26 Hz.
- On a real-world Fruit Sorting task, FreqPolicy reaches 70% success at 93.5 Hz, versus 55% for Diffusion Policy at 19.8 Hz and 60% for Flow Matching Policy at 20.2 Hz.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add frequency consistency with adaptive frequency loss (ablation) | one-step flow matching policy · Robomimic Transport | success rate: 0.84 -> 0.92 | Helped |
| add adaptive frequency loss (ablation) | full-spectrum frequency consistency · Robomimic Toolhang | success rate: 0.82 -> 0.88 | Helped |
| add spatial consistency loss (ablation) | one-step flow matching policy · Robomimic Transport | success rate: 0.84 -> 0.90 | Helped |
| add high-frequency consistency (ablation) | one-step flow matching policy · Robomimic Transport | success rate: 0.84 -> 0.88 | Helped |
| add low-frequency consistency (ablation) | one-step flow matching policy · Robomimic Transport | success rate: 0.84 -> 0.88 | Helped |
| replace selected frequency bands with full-spectrum frequency consistency (ablation) | one-step flow matching policy · Robomimic Transport | success rate: 0.90 -> 0.92 | Helped |
| replace full-spectrum consistency with adaptive frequency consistency (ablation) | FreqPolicy · Robomimic Square | success rate: 0.92 -> 0.93 | Helped |
| use frequency consistency instead of spatial-domain consistency | one-step flow policy · Robomimic Transport | success rate: 0.84 -> 0.90 | Helped |
| use frequency consistency instead of spatial-domain consistency | one-step flow policy · Robomimic Tool Hang | success rate: 0.80 -> 0.85 | Helped |
| use FreqPolicy instead of SDM | 3D visuomotor policy · MetaWorld | average success rate: 74.8% -> 78.5% | Helped |
| use FreqPolicy instead of FlowPolicy | 3D visuomotor policy · MetaWorld | average success rate: 77.2 -> 78.5 | Helped |
| use OpenVLA-FreqPolicy instead of OpenVLA-FlowMatching with 10 NFE | OpenVLA · LIBERO average | success rate: 93.7% vs. 94.8% | Helped |
Surprising findings
- Fixed low- or high-frequency-only constraints gave only minor gains, while the full-spectrum constraint and the adaptive weighting gave larger improvements.
- On the Adroit/MetaWorld table, FreqPolicy is below baselines on some tasks such as Adroit Door and MetaWorld Medium Basketball and Bin Picking.
Limitations
- Method is validated only on flow matching; extension to diffusion policies is left for future work.
- Training requires two forward passes on two random samples, which increases computation.
- VLA experiments used reduced training data (2.4M samples) compared with the original OpenVLA-OFT setting.
Experimental setup
Datasets: Robomimic Adroit MetaWorld LIBERO real-world Franka and UR demonstrations
Benchmarks: Robomimic Adroit MetaWorld LIBERO real-world Franka/UR tasks
Metrics: success rate NFE inference speed (Hz)
Baselines: DDPM DDIM RectifiedFlow ActionFlow AdaFlow Consistency Policy Consistent-FM IMLE Policy DP3 ManiCM SDM FlowPolicy
Builds on: Flow Matching Consistency-FM Consistency Policy FlowPolicy OpenVLA
Closest papers by meaning
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.94)
- FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.90)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
- Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation (NeurIPS 2026 · similarity 0.89)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.