Autonomous vehicles (AVs) have the potential to revolutionize transportation by enhancing safety, efficiency, and accessibility. A fundamental component of AV systems is the control layer, which translates high-level decisions into precise driving actions such as steering, throttle, and braking. Traditional control methods often require extensive manual tuning and perform poorly in complex or unfamiliar environments due to their limited adaptability and reliance on accurate system models. These limitations underscore the need for more flexible, data-driven approaches to AV control. This study proposes a hybrid learning-based framework that integrates imitation learning (IL) with the Soft Actor-Critic (SAC) reinforcement learning algorithm, further enhanced by a Variational Autoencoder (VAE) for efficient visual feature extraction. The framework begins with offline imitation learning using expert demonstrations to provide a robust policy initialization. This pre-trained policy is then fine-tuned through SAC in the CARLA simulator using a carefully designed reward function that promotes speed control, lane adherence, and safe navigation. The VAE encodes raw camera images into compact latent representations, improving the learning efficiency and generalization of the control policy. Experimental results demonstrate that the combined VAE-IL-SAC approach significantly improves performance.