This paper presents a unified simulation framework for training and evaluating Deep Reinforcement Learning (DRL) agents in realistic automated parking scenarios. Built entirely within the CARLA simulator, the environment includes custom scenes for perpendicular, skewed, and parallel parking layouts, designed to capture the specific spatial challenges associated with common real-world geometries. A modular Gymnasium-compatible interface exposes these scenes to standard DRL pipelines, facilitating training and experimentation. We begin by validating a redesigned 90° perpendicular scenario using a pre-trained parking policy, which retains its performance without retraining. However, when deployed in the never seen skewed and parallel layouts, the same policy fails to generalize effectively, revealing the need for targeted adaptation. To address this, we adopt a Curriculum Learning (CL) strategy that leverages prior knowledge to accelerate learning in the new introduced environments. Fine-tuned agents succeed in mastering the additional scenarios while demonstrating realistic behaviors, such as reverse maneuvers and correct alignment, even under randomized initial conditions and presence of static Non-Player Vehicle (NPV). The proposed framework aims to facilitate research in learning-based low-speed maneuvering by providing a flexible and extensible training and testing environment.