The use of Deep Learning algorithms in the domain of Decision Making for Autonomous Vehicles has garnered significant attention in the literature in recent years, showcasing considerable potential. Nevertheless, most of the solutions proposed by the scientific community encounter difficulties in real-world applications. This paper aims to provide a realistic implementation of a hybrid Decision Making module in an Autonomous Driving stack, integrating the learning capabilities from the experience of Deep Reinforcement Learning algorithms and the reliability of classical methodologies. Our Decision Making system is in charge of generating steering and velocity signals using the HD map information and sensors pre-processed data. This work encompasses the implementation of concatenated scenarios in simulated environments, and the integration of Autonomous Driving modules. Specifically, the authors address the Decision Making problem by employing a Partially Observable Markov Decision Process formulation and offer a solution through the use of Deep Reinforcement Learning algorithms. Furthermore, an additional control module to execute the decisions in a safe and comfortable way through a hybrid architecture is presented. The proposed architecture is validated in the CARLA simulator by navigating through multiple concatenated scenarios, outperforming the CARLA Autopilot in terms of completion time, while ensuring both safety and comfort.

Advances in sensor fusion techniques are redefining the landscape of 3D point cloud semantic segmentation, particularly for autonomous driving applications. We propose an enhanced approach that leverages the complementary strengths of LiDAR and multi-camera systems. This study introduces two extensions to the state-of-the-art SalsaNext model [1] based only in LiDAR: SalsaNext+RGB, which integrates RGB data into range-view (RV) images, and SalsaNext+PANO, incorporating panoramic images built from multi-camera setups. The proposed methods are evaluated using the SemanticKITTI and Panoptic nuScenes datasets, showing notable improvements in segmentation accuracy. Results indicate that RGB fusion boosts performance with minimal latency, while panoramic integration offers additional gains at the expense of higher computational load. Comparative analyses highlight significant mIoU gains, demonstrating the potential of multimodal sensor fusion for intricate driving scene understanding.

The use of DL algorithms in the domain of DM for AVs has garnered significant attention in the literature in the last years, showcasing considerable potential. Nevertheless, most of the solutions proposed by the scientific community encounter difficulties in real-world applications. This paper aims to provide a realistic implementation of a hybrid DM module in an AD stack, integrating the learning capabilities from the experience of DRL algorithms and the reliability of classical methodologies. This work encompasses the implementation of concatenated scenarios in simulated environments, and the integration of AD modules. Specifically, the authors address the DM problem by employing a POMDP formulation and offer a solution through the use of DRL algorithms. Furthermore, an additional control module to execute the decisions in a safe and comfortable way through a hybrid architecture is presented. The proposed architecture is validated in the CARLA simulator by navigating through multiple concatenated scenarios, outperforming the CARLA Autopilot in terms of completion time, while ensuring both safety and comfort.

Decision-making is a fundamental component of autonomous driving, where complex urban scenarios require safe, robust, and adaptable behaviours. This work presents a curriculum learning approach to reduce the dynamics-related reality gap in autonomous driving decision-making through a hybrid architecture that combines learning-based tactical decisions with classical planning and control methods. The proposed methodology follows a staged sim-to-real process: first, the decision-making policies are trained in a lightweight simulator to learn basic kinematic behaviours; then, they are transferred and refined in CARLA to account for vehicle dynamics; subsequently, a digital twin of the real platform and test environment is used for scenario-specific fine-tuning; finally, the resulting architecture is validated through parallel execution with a real vehicle. The proposed approach focuses on vehicle dynamics, actuation response, and scenario geometry rather than on the complete sim-to-real problem for autonomous driving. The approach is evaluated across multiple urban driving scenarios in simulation, including lane changing, roundabouts, merging, and crossroads, while real-world validation is conducted in a controlled merge scenario. Experimental results show that the proposed curriculum improves training efficiency and final performance across the different stages, achieving success rates above 91% in SMARTS. In CARLA, the proposed architecture completes the evaluated scenarios up to 50% faster than the Autopilot baseline while improving comfort and safety-related metrics in terms of acceleration and jerk. The real-world parallel execution experiment further demonstrates the feasibility of transferring the decision-making architecture to a physical vehicle under controlled conditions. Finally, an ablation study quantifies the contribution of each curriculum stage to the overall system performance.

The implementation of safe and reliable Autonomous Driving Stacks in complex urban environments remains a formidable engineering challenge. While classical modular pipelines provide necessary component-level interpretability, they are inherently rigid, often struggling to adapt to novel environments and failing to provide robust scene interpretation in highly interactive scenarios. In this paper, we present a modular End-to-End ROS-based autonomous driving architecture that upgrades a classical modular baseline by injecting learning-based models into its individual processing layers, integrating GaussianCaR and CLIP for dense semantic BEV perception, expanding the Hierarchical Petri Net state space for safe multi-agent reasoning, refining the planning layer with continuous curve optimization, and replacing the previous reactive controller with an Adaptive Nonlinear Model Predictive Control strategy for superior trajectory tracking. Validated in the CARLA simulator across challenging traffic scenarios and adverse environmental conditions, the proposed architecture raises the Driving Score from 53.81% to 66.46% over the previous baseline, driven by a substantial increase in the Infraction Penalty from 0.59 to 0.79, reflecting a fundamental shift towards safer and more conservative driving behavior at the cost of a moderate reduction in route completion. Against pure End-to-End approaches, our architecture achieves the highest Driving Score at 73.9% and the strongest Infraction Penalty at 0.913, demonstrating that modular interpretability and competitive End-to-End performance are not mutually exclusive. Code will be made publicly available online.