AT-Drive: Exploiting Adversarial Transfer for End-to-end Autonomous Driving

Dapeng Zhang, Zhenlong Yuan, Kaer Huang, Yiqiang Yan, Chenyang Li, Hongtao Nie, Shiyue Zhao, Rui Zhou, Qingguo Zhou

Keywords: CV, Applications, 3D vision, Robots, Autonomous Driving

TL;DR: AT-Drive is an approach that transfer the simulated imitation driving capabilities to real-world deployment.

Abstract: End-to-end autonomous driving methods aim to enable robust vehicle control by imitating successful driving behavior. Existing approaches are trained either on real-world data, which closely reflects practical applications, or on simulation data, which is used to simulate undesirable behaviors such as non-compliant driving habits, car accidents, and off-road scenarios. However, these methods fail to integrate the advantages of both data sources effectively. In this paper, we propose AT-Drive, an end-to-end adversarial transfer framework for autonomous driving. AT-Drive is the first approach that transfer the simulated imitation driving capabilities to real-world deployment. AT-Drive first pre-trains simulation and real-world model with simulation and real-world dataset separately. Then, two discriminators are utilized to adversarially train the real-world model, producing a model that transfers simulation-based driving capabilities into real-world deployment. This approach bridges the gap between simulation and real-world autonomous driving. Furthermore, by incorporating a unique back-propagation strategy, AT-Drive achieves state-of-the-art performance on the newly partitioned nuScenes dataset.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 1438