Abstract
<title>Abstract</title> <p>Autonomous vehicle navigation relies on multi-sensor fusion combining IMU, GPS, and radar for continuous state estimation. Two fundamental limitations persist in existing approaches: Kalman filter noise parameters are fixed at design time and cannot adapt to changing driving conditions, and motion classification operates independently from state estimation with no information shared between them. This paper proposes a regime-aware hybrid multi-sensor fusion framework that addresses both limitations through a unified two-phase pipeline tested on the nuScenes v1.0-mini real-world driving dataset. In Phase~1, statistical features extracted from IMU, GPS, and radar are ranked via Gradient Boosting importance scoring, binarized, and classified into three driving regimes- stopped, cruising, and turning using a discrete Hopfield neural network with Hebbian learning, alongside Gradient Boosting, Random Forest, and a Neuromorphic Fusion classifier. In Phase~2, the classified regime label is passed as an eleventh dimension of the reinforcement learning agent state vector, making the agent regime-aware as it adaptively tunes the noise parameters of three Kalman filter variants- EKF, UKF, and square-root CKF using TD3, SAC, PPO, and an Ensemble RL agent. A GPS spoofing defense integrating Normalized Innovation Squared monitoring, CUSUM change detection, and a parallel trusted estimator with state rollback is incorporated within the same pipeline. Performance is evaluated across multiple configurations and independent random seeds using position RMSE, with statistical significance verified through paired Wilcoxon signed-rank tests.</p>