This paper propose a latent flow matching model using time-conditioned Fourier layers to capture multi-scale modes with high fidelity in both spatial and spectral characteristics to forecast high-dimensional, PDE-governed dynamics over long time horizons.

Something interesting I learned from this paper:

  1. Flow matching can be used for timeseries forecasting
  2. Flow matching can be used on the top of latent space

Since robotic trajectories are also timeseries, and contains both spatial and spectral characteristics, and high-dimensional. Can we use TEMPO in trajectory generation?