UC San Diego

The sixth-generation (6G) cellular ecosystem is evolving toward integrated sensing and communication (ISAC), where the same infrastructure supports both high-rate data exchange and environmental perception. Vehicular networks are among the most demanding use cases: autonomous navigation demands ultra-reliable connectivity and centimeter-level positioning accuracy, yet conventional beam training protocols incur prohibitive overhead and global navigation satellite system (GNSS)-based localization fails in urban canyons. This dissertation develops a hybrid model/data-driven framework addressing these challenges through three contributions. The first introduces a passive radar-aided framework for multiuser millimeter wave (mmWave) link configuration. A roadside unit senses existing automotive frequency modulated continuous wave (FMCW) transmissions from multiple vehicles via a mixing filter bank with constant false alarm rate (CFAR) detection, isolating individual vehicle signals from multiuser interference. Three deep neural network architectures then map estimated radar spatial covariances to communication covariances, compensating for frequency and geometry mismatches between the 76 GHz radar and 73 GHz communication bands. The covariance-prediction-assisted approach reduces beam search space by up to 32 times and improves the achievable rate by 21.9% over radar-only methods. The second addresses high-accuracy 3D vehicle localization from a single base station snapshot. Two time-domain channel estimation algorithms, a two-stage multidimensional orthogonal matching pursuit (MOMP) variant and ESPRIT-D (an off-grid subspace method), extract multipath parameters while accounting for system filtering effects and unknown clock offsets. A lightweight network, PathNet, classifies estimated paths to select geometrically useful components, while a Transformer-based ChanFormer refines initial geometric position estimates through cross-attention, achieving 28 cm accuracy for 80% of users in line-of-sight (LOS) and sub-meter accuracy for 55% in non-line-of-sight (NLOS). The third extends single-snapshot localization to continuous tracking and collaborative multi-vehicle positioning. F-MOMP enables efficient high-resolution channel tracking by exploiting temporal correlation and factored dictionaries. VO-ChAT and VP-ChAT leverage channel sequence history through spatial-temporal and cross-attention mechanisms to track vehicle orientation and refine position estimates. For NLOS vehicles, a model-based collaborative framework exploits vehicle-to-vehicle (V2V) sidelinks to geometrically estimate per-vehicle clock and orientation offsets, achieving 0.3 m average error with sub-meter accuracy for 90% of cases. Together, these contributions demonstrate that embedding physical-layer signal models into deep learning architectures yields ISAC systems that are accurate, computationally efficient, and interpretable, key properties for safety-critical vehicular deployments.