| Summary: | Autonomous vehicles (AVs) operate in highly dynamic environments, making them essential in modern transportation systems. Their success depends on real-time decision-making in unpredictable traffic scenarios, which are often beyond the scope of initial design assumptions. This unpredictability limits the effectiveness of traditional rule-based decision-making systems and predefined cost functions for real-time optimization. In critical applications like autonomous driving, reinforcement learning (RL) agents without safety mechanisms often struggle to converge or require extensive training data to produce reliable policies, leading to challenges in achieving safe and efficient operation. To address these challenges, this paper proposes a reinforcement learning (RL)-based framework. The ego vehicle refines its decision-making abilities by interacting with a simulated traffic environment. A short-horizon safety mechanism (SM) is integrated to ensure safer training by providing alternative safe actions during critical scenarios. The RL agent employs a deep neural network to map system states to optimal actions. The SM generalizes risky states, such as near-misses or collisions, rainy environment during night while creating a stable learning environment that enhances learning efficiency and enables meaningful exploration for optimal policy development. The method was validated in a highway driving scenario with varying traffic densities using the DQN algorithm and the CARLA simulator. Results demonstrated that the integration of the safety mechanism significantly improved learning efficiency and enabled the AV to make safe, reliable decisions even in complex and unpredictable traffic conditions. © The 2025 International Conference on Artificial Life and Robotics (ICAROB2025).
|
|---|