RaVF: Learning Radar Velocity Fields via Spatial-Doppler Guidance

Scene-level Velocity field estimation from 4D millimeter-wave radar is critical for robust perception but remains challenged by the inherent sparsity and noise of point-based paradigms. Furthermore, Doppler ambiguity induced by the Nyquist sampling limit fundamentally disrupts the continuity of velocity field learning. To address these challenges, we propose RaVF, a physically grounded framework that directly learns dense velocity fields from high-fidelity spatial-Doppler spectra. RaVF is built upon three key designs tailored to radar-specific signal distortions. (1) To mitigate range-dependent propagation attenuation and nonlinear angular spatial quantization, we devise a Physically-aware Encoder that calibrates spectral feature extraction within the native radar beamforming. (2) Observing that ego-induced radial motion explains static Doppler responses, we factor out static motion and guide dynamic flow estimation for physically consistent velocity field learning. (3) To enhance kinematic plausibility, we introduce an ego-conditioned bidirectional pyramidal flow module that transfers backward flow into the forward frame and enforces anti-symmetric forward-backward consistency. Extensive experiments demonstrate that RaVF significantly outperforms state-of-the-art baselines on velocity fields. Our code will be available.