DL-ALP: A Deep Learning-Driven Adaptive Trajectory Planner for Low-Speed Complex Dynamic Environments

In this study, we present DL-ALP (Deep-Learn Adaptive Low-Speed Planner), a local trajectory planning framework for autonomous vehicles in low-speed, dynamic, and spatially constrained environments such as parking lots or residential-type areas. Existing methods often struggle in these environments because of unpredictable dynamic obstacles and more complex interaction patterns. A key novelty of DL-ALP is that it incorporates into a framework an environment intent perception module based on Graph Neural Networks that, in real-time, can predict the likely behaviors and risk regions of surrounding agents. The environments’ dynamic risk information is then integrated into a hybrid trajectory generation and evaluation framework, combining a small set of spline-based trajectory candidates with a scoring network based on deep reinforcement learning. Our proposed method generates trajectories that provide a reasonable compromise between safety, comfort, and interaction efficiency. Experimentation on a simulation dataset and real-world datasets provides evidence that DL-ALP has robust performance and does well in terms of adaptability to complex low-speed driving environments.