To enable autonomous driving in real-world environments that involve a diverse range of geographic variations and complex traffic regulations, it is essential to investigate Deep Reinforcement Learning (DRL) algorithms capable of policy learning in high-dimensional environments characterized by intricate state–action interactions. In particular, closed-loop experiments, which involve continuous interaction between an agent and their driving environment, serve as a critical framework for improving the practical applicability of DRL algorithms in autonomous driving systems. This study empirically analyzes the capabilities of several representative DRL algorithms—namely DDPG, SAC, TD3, PPO, TQC, and CrossQ—in handling various urban driving scenarios using the CARLA simulator within a closed-loop framework. To evaluate the adaptability of each algorithm to geographical variability and complex traffic laws, scenario-specific reward and penalty functions were carefully designed and incorporated. For a comprehensive performance assessment of the DRL algorithms, we defined several driving performance metrics, including Route Completion, Centerline Deviation Mean, Episode Reward Mean, and Success Rate, which collectively reflect the quality of the driving in terms of its completeness, stability, efficiency, and comfort. Experimental results demonstrate that TQC and SAC, both of which adopt off-policy learning and stochastic policies, achieve superior sample efficiency and learning performances. Notably, the presence of geographically variant features—such as traffic lights, intersections, and roundabouts—and their associated traffic rules within a given town pose significant challenges to driving performance, particularly in terms of Route Completion, Success Rate, and lane-keeping stability. In these challenging scenarios, the TQC algorithm achieved a Route Completion rate of 0.91, substantially outperforming the 0.23 rate observed with DDPG. This performance gap highlights the advantage of approaches like TQC and SAC, which address Q-value overestimation through statistical methods, in improving the robustness and effectiveness of autonomous driving in diverse urban environments.

This study addresses the optimization of a camera-based bird’s eye view (BEV) segmentation technique that operates in real-time within an embedded system environment while maintaining high accuracy despite limited computational resources. Specifically, it examines three technical approaches for BEV segmentation in autonomous driving: depth-based methods, MLP-based methods, and transformer-based methods, focusing on key techniques such as lift–splat–shoot, HDMapNet, and BEVFormer. A mathematical analysis of these methods is conducted, followed by a comparative performance evaluation using the nuScenes dataset. The optimization process was carried out in three stages: accuracy improvement, latency reduction, and model size optimization. In the first stage of the process, the three modules for BEV segmentation (encoder, view transformation, and decoder) were selected with the goal of maximizing mIoU performance. In the second stage, environmental variable optimization was performed through input resolution adaptation and data augmentation to improve accuracy. Finally, in the third stage, model compression was applied to minimize model size and latency for efficient deployment on embedded systems. Experimental results from the first stage show that the lift–splat–shoot view transformation model, based on the InternImage-B encoder and EfficientNet-B0 decoder, achieved the highest performance with 54.9 mIoU at an input image size of $448\times 800$. Notably, the lift–splat–shoot view transformation model with the InternImage-T encoder and EfficientNet-B0 decoder demonstrated performance of 53.1 mIoU while achieving high efficiency (51.7 ms and 159.5 MB, respectively). The application of the second stage revealed that increasing the input resolution does not always lead to improved accuracy, and there is an optimal resolution size depending on the model. In this study, the best performance was achieved with an input image size of $448\times 800$. During the third stage, FP16 quantization enabled a 50% reduction in memory size and decreased latency while maintaining similar or identical mIoU performance. When deployed on the NVIDIA AGX Orin device, which operates under power constraints, energy efficiency improved, although it resulted in higher latency under certain power supply conditions. As a result, the InternImage encoder-based lift–splat–shoot technique was shown to achieve the highest accuracy performance relative to latency and model size. This approach outperformed the original method by achieving a 29.2% higher mIoU while maintaining similar latency performance and reducing memory size by 32.2%.