DefinePK

Abstract

Machine learning (ML) algorithms play a pivotal role in the key functional areas of autonomous vehicle (AV) navigation, including perception, localization, mapping, trajectory prediction, planning, and control. Despite the advancements in sensor technologies such as high-definition cameras, LiDAR, and radar—commonly used for mapping, obstacle detection, and localization—autonomous vehicles still face significant challenges in reliably navigating unfamiliar environments with unpredictable dynamics. These challenges stem from real-world factors such as weather variability, traffic congestion, pedestrian activity, and erratic behaviour of other drivers. Machine learning offers a promising solution to address these complexities. As one of the most rapidly evolving technologies, ML enables autonomous navigation systems to better interpret sensory data, adapt to dynamic surroundings, and make informed driving decisions. In addition to enhancing traffic safety and minimizing human error-related accidents, connected and autonomous vehicles (CAVs) can also fulfil a wide range of smart functions—from last-mile delivery services to urban surveillance in smart cities. To achieve these benefits, autonomous vehicles must be capable of independently reaching their destinations while cooperating with road infrastructure. Recent advancements in Cooperative Vehicle-Infrastructure Systems (CVIS) facilitate seamless communication between AVs and elements like traffic lights (TLs), promoting safer and more efficient transportation systems. CVIS allows for the integration of real-time information sharing between infrastructure and vehicles, thus supporting more accurate navigation and situational awareness. This study focuses on evaluating the practicality of two well-known reinforcement learning techniques—Proximal Policy Optimization (PPO) and Deep Q-Network (DQN)—within autonomous navigation systems. The assessment began with training the models in a low-fidelity driving simulator, followed by testing in a high-fidelity traffic simulation environment to replicate more realistic driving conditions. Multiple driving scenarios were considered to evaluate the robustness, adaptability, and performance of each algorithm. Results indicate that both PPO and DQN outperform traditional models, with PPO showing superior performance in maintaining consistent speed, navigating efficiently, and minimizing idle or ineffective movements. In autonomous driving, vehicles must constantly evaluate the state of surrounding objects, whether static or in motion, and adapt their behaviour accordingly. To support this, machine learning techniques such as convolutional neural networks (CNNs) optimized with Adaptive Moment Estimation (Adam), and neuro-fuzzy systems fine-tuned through Particle Swarm Optimization (PSO), are employed. These methods enable real-time decision-making and smooth control, empowering AVs to respond swiftly and accurately based on learned patterns from extensive training data