Hybrid Reinforcement Learning-Based Collision Avoidance Algorithm for Autonomous Vehicle Clusters

Nowadays, collaborative collision avoidance for autonomous vehicle clusters has become the key to ensure traffic safety. Aiming at the complex and changeable traffic environment, this paper proposes a novel collision avoidance method for Autonomous Vehicle Clusters based on hybrid reinforcement learning. The method combines the adaptive capability of reinforcement learning with the feature extraction capability of deep learning to improve the collision avoidance performance in complex traffic scenarios. A hybrid reinforcement learning framework is designed, which consists of a deep neural network structure and a reinforcement learning structure. The feature extraction key from environmental perception data and predict possible collision risks. While the latter learns how to adjust vehicle motion parameters based on these features and the historical performance of collision avoidance strategies. Convolutional neural network is used to process image data to capture spatial information in traffic scenes, and time series data is combined with long and short time memory network to capture the time dependence of vehicle motion. Through a large number of simulation experiments and field tests from KITTI (Karlsruhe Institute of Technology and Toyota Technological Institute), nuScenes and INTERACTION, we verify the effectiveness of the proposed method. The experimental results show that the proposed algorithm in this study showed a high success rate of collision avoidance and a low average reaction time under most collision avoidance times.