Online Imitation Learning for Self-Driving Simulation

Zhe Zhang, Sanyuan Zhao

Abstract: The end-to-end autonomous driving policy has made great progress with the development of deep learning. The current methods are mainly divided into imitation learning and reinforcement learning. The method of imitation learning can quickly realize the one-to-one correspondence between states and actions, but is limited by the dataset and is prone to overfitting. Therefore, the current methods mainly focus on extracting more robust input state features and proposing a more generalized dataset. Reinforcement learning methods can obtain richer input states due to online training, but at the same time requires longer training time, so current methods mainly focus on reducing training time and designing appropriate rewards. In this paper, we propose an end-to-end temporal convolution model based on segmentation medium, which uses online imitation learning to obtain richer input states, train more robust policy networks. At the same time, to reduce the training time, we use our own designed segmentation medium to replace the raw sensor information as the input of the policy network. Experiments on the CARLA driving benchmarks show that our approach achieves satisfactory results and has excellent generalization ability.