Optimization and Safety Control of Autonomous Driving Strategy Based on Adaptive Reward Signal

Engineering notes

The experimental results show that the proposed method outperforms traditional methods under different traffic densities. In low-density scenes, the average speed of AR-E-DQN is 28.52 m/s, the average reward is 48.63, and the collision rate is only 1.2%, significantly better than DQN's 25.86 m/s, 37.63, and 5.2%.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

In order to improve the efficiency and safety of autonomous driving in complex traffic environments, this study proposes a hierarchical control framework based on Adaptive Reward Signal (AR-E-DQN) and Adaptive Responsibility Sensitive Safety Model (ARSS). The experimental results show that the proposed method outperforms traditional methods under different traffic densities. In low-density scenes, the average speed of AR-E-DQN is 28.52 m/s, the average reward is 48.63, and the collision rate is only 1.2%, significantly better than DQN's 25.86 m/s, 37.63, and 5.2%. In high-density scenes (ρ = 2), adaptive rewards can still maintain a speed of 24.76 m/s and a low collision rate of 5.3%, while traditional fixed rewards can reduce the speed to 20.32 m/s and increase the collision rate to 12.8%. Similarly, ARSS significantly outperforms traditional RSS in terms of safety control, reducing collision rates by over 50% at high densities. Research has shown that this method achieves a better balance between efficiency and safety, and has strong practicality.

Links and sources

Comments