Risk-Aware Attention-Based TD3 Approach for Vehicle Decision-Making in Dynamic Traffic

It is of paramount importance to consider driving safety when developing autonomous vehicles. However, the decision-making models for autonomous driving (AD) face significant challenges, including high scene interactivity, high collision probability, and difficulty in guaranteeing decision safety. To guarantee driving safety, we propose a vehicle behavior decision-making framework based on a risk-aware attention-based TD3 (RAB-TD3) approach, which aims to identify an optimal driving decision-making strategy that ensures efficiency while minimizing the expected risk. Firstly, an uncertainty-based benefit function is constructed for the assessment of risk in a vehicle's environment which generates a safety evaluation index based on information such as distance, speed, and uncertainty. Secondly, the self-attention mechanism is introduced into the evaluation of environmental state information to consider the key interacting vehicles in the decision-making process. Ultimately, a barrier function is constructed based on lane change safety information, which is then combined with a risk-benefit function to direct vehicles to select their maneuvers in a manner that prioritizes safety. The simulation results in the Highway-Env environment show the Approach reduces the unsafe behavior ratio to below 0.05. Additionally, the model improves learning convergence speed and vehicle passing efficiency compared to traditional DDPG, SAC and TD3 algorithms.