Collision avoidance is considered one of the essential features in Autonomous Vehicle (AV) applications. In this research, we suggest combining the DQN with the new real-world concept called “Quarrel” ( \(DQ^{2} N\) ). This hybrid framework leverages the strength of the Multiagent Reinforcement Learning (MARL) concept, which allows two AVs, one acts as a pursuer and the other as an evader, to learn the best strategies to interact in a competitive adversarial environment. The main goal is to ensure the safety of these agents by predicting and responding to each agent’s actions in real time and guaranteeing effective collision avoidance. The word “quarrel” refers to the competitive, opposing interaction between agents (evader and pursuer) who aim to maximise their own goal by reacting against the other’s activities. The pursuer seeks to catch or overtake the evader, whereas the evaders aim to break the pursuer’s capture motive. Here, the novel Quarrel method combines the main principles of game theoretic and MARL approaches. The Quarrel mechanism enriches the decision-making process and reward structure with competitive strategies by augmenting the MARL algorithms. This guarantees that agents in an adversarial pursuit change their strategy depending on the actions of their opponents while maximising their aim. In pursuit–evasion situations, the quarrel idea improves the agents’ capacity to predict or observe the other agent’s behaviours very effectively. This proposed approach is tested in both simulated and real-world environments. To simulate the proposed approach, we follow the study procedures. Similarly, in real-world situations, we create the AV racing app to observe the efficiency of two agents and their decision-making ability under real-world applications. The results show considerable improvements in collision avoidance, overtaking strategies, and overall system efficiency.
