RadarDreamer: Generative Radar World Model from Instantaneous Doppler

Sensor-level world models anticipate future observations for predictive perception, planning, data synthesis, and closed-loop simulation, yet their development has largely centered on visual, occupancy, and geometric spaces. Unlike passive visual observations, a single radar measurement already encodes both spatial structure and instantaneous radial motion, making native radar particularly informative for future observation generation. The challenge is twofold: RA and DA retain spatial and Doppler evidence only as shared-azimuth marginals, obscuring their underlying structure–motion association. Meanwhile, future echoes entangle predictable motion evolution with stochastic variations in visibility, reflectivity, multipath, and measurement noise. Our key insight is to exploit instantaneous Doppler as structured dynamics evidence for stochastic future generation. We present RadarDreamer, a conditional rectified-flow model that maps Gaussian radar videos to five coherent future RA+DA pairs. Exact multiband states preserve measurement coefficients, azimuth-indexed coupling exchanges spatial and Doppler evidence while retaining their axis semantics, and radar-aware objectives constrain sparse innovation, temporal evolution, response structure, and projection consistency. On ColoRadar, spatial–Doppler coupling reduces CRPS by 2.16% and improves expected-sample innovation F1 by 10.98% over an architecture-matched generator, while a single RA+DA observation reduces joint log-MAE by 5.86% relative to five-frame RA history. Together with four-domain rollouts, these results demonstrate improved motion fidelity and measurement realism, establishing conditional native-radar generation as a predictive sensing primitive for radar-centric world models.