中国科学院行为科学重点实验室知识库

Traditional navigation models in autonomous driving rely heavily on metric maps, which severely limits their application in large scale environments. In this paper, we introduce a two-level navigation architecture that contains a topological-metric memory structure and a deep image-based controller. The hybrid memory extracts visual features at each location point with a deep convolutional neural network, and stores information about local driving commands at each location point based on metric information estimated from egomotion information. The topological-metric memory is seamlessly integrated with a conditional imitation learning controller through the navigational commands that drives the vehicle between different vertices without collision. We test the whole system in teach-and-repeat experiments in an urban driving simulator. Results show that after being trained in a separate environment, the system could quickly adapt to novel environments with a single teach trial and follow route successively under various illumination and weather conditions.