CORTEX: Cooperative Occlusion- Resilient Trajectory Execution via Request-Aware V2I Fusion

Abstract

Traditional end-to-end learning architectures for autonomous driving hit a strict physical limitation in dense urban environments due to severe line-of-sight occlusion, inducing elevated aleatoric uncertainty and passive freezing behaviors at intersections. To resolve this perception bottleneck, we introduce CORTEX: Cooperative Occlusion-Resilient Trajectory Execution via Request-Aware V2I Fusion. CORTEX resolves occlusion-induced perceptual collapse by coupling infrastructure-expanded line-of-sight awareness with task-driven feature optimization. Elevated Roadside Units (RSUs) expose concealed blind-zone obstacles, while a request-aware hard-masked cross-attention bottleneck queries exclusively along the predicted driving corridor to eliminate redundant background bandwidth. Concurrently, an explicit kinematic latency corrector applies affine grid warping to neutralize asynchronous transmission lags under deterministic step delays up to 500 milliseconds. Furthermore, a topography-preserving spatial convolutional control head maps the fused spatio-temporal features directly onto physical actuators within a non-truncated 13,960-channel control input space. We evaluated this unified pipeline offline on the CARLA Town05 validation split of the V2XVerse dataset. The empirical results confirm that CORTEX achieves a global average displacement error of 0.3157 meters, outperforming standalone baselines by over 46% while securing 70.4% volumetric network traffic relief on the left-turn corridor and 45.1% under dense occluded right-turn requests (with an average 61.1% overall data volume reduction) and exhibiting inherent structural occlusion resilience against localization pose noise. The source code and pre-trained models are publicly available at: https://github.com/hamid79daneshvar/CORTEX