Bidirectional neural network for trajectory planning: An application to medical emergency vehicle

This paper focuses on the autonomous trajectory optimization of in-hospital emergency vehicles, aiming to plan a fast, safe, and comfortable trajectory for real time navigation to reach the emergency room. This has direct and significant practical implications for saving patients’ time and reducing the burden on medical staff. To address it, we propose a novel real-time trajectory planning algorithm and introduce a training framework that contributes to planner's transferability. In the dataset preparation phase, we employ an optimization-based method to solve the in-hospital trajectory planning problem, generating substantial trajectories. In the training phase, we design a neural network-based planner to establish logical connections between states and control commands. To enhance interaction with the vehicle itself, we design a novel network training framework that incorporates a neural network-based vehicle simulator to learn the vehicle's self-information, facilitating training the planner. During the inference phase, our planner can plan a collision-free, time-efficient, and comfortable trajectory. Furthermore, our algorithm demonstrates ease of transferability to different vehicle models. Finally, extensive simulations experiments are conducted to validate the safety, speed, and comfortability of the algorithm's trajectory planning, as well as its excellent transferability.