Towards context-aware trajectory planning for autonomous vehicles

Abstract

Autonomous driving is an emerging technology with significant potential to improve road safety, reduce traffic congestion, and enhance transportation efficiency. Despite significant progress in recent years, planning a safe, comfortable, efficient, and traffic-aware trajectory for autonomous vehicles in the real world remains challenging. To address this issue, the dissertation proposes making planners aware of the autonomous vehicle's neighborhood context to ensure proper navigation on the road through traffic without causing any collisions or traffic jams. Our first proposed approach is a decentralized multi-agent trajectory planning algorithm that adjusts responsibility distributions between agent pairs based on their common neighbors' information. In addition, we present a new trajectory prediction technique that uses occupancy grid maps to reduce unrealistic predictions and improve safety. Furthermore, to directly learn a planning policy from human driver demonstrations, we introduce a context-conditioned imitation learning method that addresses the covariate shift issue in offline imitation learning. Finally, we develop a long-term microscopic traffic simulator to safely test the autonomous driving policy and assess its impact on overall traffic efficiency. All the proposed algorithms have been validated in simulated environments, and the dissertation concludes with a discussion on the potential of translating these techniques into real-world traffic systems in the near future.