Vehicular Edge Computing (VEC) has improved significantly, offering high-bandwidth and low-latency services for connected autonomous vehicles. VEC is particularly useful for applications such as autonomous driving, augmented reality, and vehicular safety systems that require low latency. However, since vehicular networks are dynamic and computational resources are inconsistent, intelligent decision-making is necessary for effective task offloading. Task offloading is a process that distributes computation closer to the point of data generation, reducing communication latency and enabling vehicular environments to process data and make real-time decisions. Therefore, managing computational resources through task-offloading decisions is crucial for seamless data processing and real-time applications. This study proposes a machine learning (ML)-based task offloading architecture for CAVs edge computing networks. First, a comprehensive taxonomy of ML-driven task offloading techniques for smart vehicular edge computing is presented. Then, existing ML-based techniques are classified under supervised, unsupervised, reinforcement, and deep learning-based schemes. Additionally, edge computing simulators, workload benchmarks, and performance metrics have been compared. Finally, research challenges and potential opportunities for future trends in edge-computing-based task offloading schemes have been discussed.