Research map / Reinforcement learning / Visual and representation RL

Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer · 0 citations (OpenAlex)

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 Frequency-Aware Flow Matching (FAFM), which runs flow matching over DCT coefficients of action chunks and adds a first-order derivative loss, yielding continuous, temporally smooth actions for standalone policies and VLAs.

Problem

Action-chunk flow-matching policies discretize trajectories, which makes them brittle to demonstrations at heterogeneous control frequencies and prone to temporally inconsistent, jittery actions.

Approach

Perform flow matching in a DCT cosine-basis coefficient space anchored to physical time so that trajectories at different frequencies map to the same targets, reconstruct continuous actions analytically, and add a derivative-matching loss that equals a weighted H^1 Sobolev penalty without extra parameters.

Key results

• On LapGym, FAFM reaches 97% SR on RT versus 94% for the best baseline (MPD), and -7.57 LDLJ vs -9.38 for MPD.
• On mixed-frequency LIBERO drawer, pi0 drops from 94% (single frequency) to 0% (mixed), while FAFM keeps 92% under both conditions.
• On real-robot pi0.5 tasks, FAFM gains +15% SR on pick-and-place and +10% on obstacle avoidance over the pi0.5 backbone.

Surprising findings

• Standard pi0 collapses from 94% to 0% success on a mixed-frequency version of the LIBERO drawer task with the same number of demonstrations.

Limitations

• Less effective for tasks needing high-frequency or impulsive actions due to DCT truncation.
• Many results are on small-scale or single-task settings (toy, obstacle avoidance, 20-episode real-robot evaluations).
• Real-robot results rely on 20 episodes per method.

Experimental setup

Datasets: LIBERO LapGym synthetic toy bimodal trajectories PyBullet obstacle avoidance demonstrations

Benchmarks: LIBERO LapGym synthetic toy bimodal benchmark PyBullet obstacle avoidance real-world Franka pick-and-place

Metrics: success rate LDLJ number of trajectory modes

Baselines: Diffusion Policy Flow Matching Policy Streaming Flow Policy FreqPolicy MPD pi0 pi0.5

Builds on: flow matching pi0 FAST Diffusion Policy MPD

Closest papers by meaning

• FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.91)
• FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.89)
• Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.88)
• ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation (CVPR 2025 · similarity 0.87)
• Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.86)