Parallel Kalman Filtering with Physics-Informed Selective State-Space Models for Robust Radar Sensing

We propose a physics-informed long-sequence Kalman filtering framework that leverages state-space neural architectures to model extended temporal structure beyond the limits of frame-by-frame filtering. By incorporating known system and observation matrices, the model learns an adaptive Kalman gain with substantially reduced training requirements. Evaluated on the RADIATE dataset, the proposed method achieves 5.5 dB higher accuracy under clean conditions, delivers 9.5 dB improvement under 1% frame-drop scenarios, and provides 2.4x faster sequence-level inference than neural-network-assisted Kalman filters.