Robust MPC-based Trajectory Planning for Autonomous Driving under Occlusion

Abstract

Abstract With the rapid advancement and widespread adoption of autonomous driving, trajectory planning must not only ensure efficiency and comfort but also guarantee safety. Planning under occlusions, however, remains a longstanding challenge. In situations such as sudden pedestrian emergence, conventional methods based on sampling or optimization primarily account for visible obstacles, while end-to-end deep learning approaches, although capable of implicitly considering occlusion effects, often suffer from black-box characteristics and limited interpretability. To address this issue, we propose a lightweight occlusion-aware robust MPC trajectory planning module. The module can be seamlessly integrated into existing planners and is selectively activated in high-risk occlusion scenarios to enhance safety. Using reachable set analysis, we explicitly model hidden road users and their motion predictions, which are incorporated into a carefully designed Robust Model Predictive Control (RMPC) framework. Two representative cases, hidden pedestrians and hidden vehicles, are investigated through extensive simulation studies. Compared with a baseline MPC, our approach reduces the collision rate from over 10% to 0%, demonstrating its effectiveness in ensuring safe navigation under occlusions.