Learning-based trajectory planning for AGVs in dynamic environment

In this work, we present a learning-based framework for rapid trajectory planning of autonomous ground vehicles (AGVs) in dynamic environments. The approach integrates optimization techniques with deep learning to design a real-time planner capable of generating kinematically feasible trajectories. A continuous iterative method is first developed for dataset construction, enabling efficient generation of optimal trajectory sets. Based on this dataset, a neural network is trained to learn the mapping between AGV states and actions while capturing their temporal dependencies. During online planning, the trained model produces decision actions from the current state and sensor feedback, enabling real-time planning of safe and feasible trajectories. Results demonstrate the effectiveness of the proposed framework.