Collaborative and safe autonomous driving through multi-agent deep reinforcement learning

Autonomous Vehicles (AVs) have the potential to drastically improve traffic safety and efficiency by eliminating human error. Despite substantial advancements, fully autonomous deployment faces challenges in navigating unpredictable urban environments, ensuring cybersecurity, and coordinating multiple AVs in real-time. Traditional Reinforcement Learning methods succeed in isolated settings but fall short in multi-agent scenarios. This thesis uses Multi-Agent Reinforcement Learning (MARL) to improve AV decision-making in dynamic, real-world conditions. This research develops secure MARL solutions for AV deployment, addressing coordination, communication, security, scalability, and sim-to-real challenges. Key contributions include a domain randomisation technique developed to enhance the transferability of simulation-trained MARL policies to real-world applications. A novel method for AV coordination in emergency scenarios using MARL together with risk-based modelling. Advanced communication strategies facilitate AV navigation in complex urban environments, particularly in occluded scenarios. A robust defence framework to protect against adversarial communication attacks, ensuring reliable network cooperation. Finally, a curriculum-based MARL approach is proposed for large-scale autonomous systems, demonstrating learning stability and efficiency. Experimental results demonstrate the effectiveness of these contributions. The AV coordination method for emergency vehicles reduces collision risks and enables 15% higher speeds while maintaining high safety distances. Communication-based policies with LiDAR feature sharing reduce crash rates by 70%, and a communication scheduler decreases communication efforts by 60%. The defence mechanism shows zero-shot resilience in most adversarial scenarios, suffering 50% less disruption and quickly restoring cooperative performance. The scalable method for large-scale systems improves total reward by 31% while accelerating learning speed by 47%. Transferring trained policies with domain randomisation reduces the reality gap by 90%. These advancements collectively enhance AV capabilities in coordination, scalability, security, and practical applicability. However, challenges remain in applying these solutions in real-world scenarios, such as addressing communication constraints, navigating regulatory and privacy concerns, and further bridging the reality gap for real-world scenarios.