Soft collision avoidance based car following algorithm for autonomous driving with reinforcement learning

Abstract

By safety supervision on dangerous driving behaviors, emergent braking in autonomous vehicles can effectively prevent collisions when using the car following algorithm based on deep reinforcement learning. However, the significant deceleration associated with emergent braking often results in an uncomfortable driving experience and high energy consumption. To address this issue, a soft collision avoidance based car following algorithm is proposed. Different from emergent braking, our approach introduces a deceleration adjustment value to the current acceleration output. This adjustment value is calculated by considering safe distance with attenuation coefficient in terms of multi-step prediction, while the attenuation coefficient and the predicted time step are discussed in detail. Comparative analysis, including statistical results and representative cases, demonstrates that the proposed algorithm significantly enhances driving comfort (improve 37.341 %) and reduces energy consumption (improve 11.244 %) without increasing collision risks.