Brief

Researchers have developed Frequency-Aware Flow Matching (FAFM), a novel method for generating continuous and temporally consistent robotic actions. This approach addresses limitations in existing flow matching techniques that struggle with heterogeneous control frequencies and can produce unstable actions. FAFM transforms actions into the frequency domain using discrete cosine transform, performs flow matching on these coefficients, and reconstructs continuous actions. It also incorporates a temporal derivative regularization to ensure smooth movements. The method has demonstrated improvements in success rates, expressivity, and robustness across various benchmarks and on a real-world robot. AI