Multimodal Sensor Fusion Optimization through Post-Training for Autonomous Vehicles

Abstract

Multimodal sensor fusion is central to autonomous vehicle perception because no single sensing modality can provide reliable geometry, semantics, and resilience across weather, lighting, and traffic conditions. Although architectural advances have improved fusion, deployed systems remain constrained by distribution shift, calibration drift, sensor degradation, and long-tail events. This paper examines post-training as a system-level optimization layer for multimodal sensor fusion. Post-training includes supervised fine-tuning, parameter-efficient adaptation, uncertainty calibration, knowledge distillation, and preference or reward-based alignment. The argument is that these techniques should not be treated as model-centric refinements but as infrastructure-aware mechanisms that shape the trade-offs among latency, redundancy, robustness, energy, and regulatory accountability. The paper analyzes fusion architectures, describes how post-training alters representation alignment and failure behavior, and considers deployment constraints in edge and vehicle-cloud environments. It further discusses governance implications, including validation, auditing, fairness, and sustainable mobility. The conclusion is that post-training can improve robustness and adaptation, but its benefits depend on data governance, sensor health monitoring, uncertainty communication, and institutional capacity to certify continuously updated systems.