Data-driven Motion Planner for Autonomous Vehicle using an Ensemble of Global Map and Local view

We present a data-driven motion planner for the trajectory generation in a complex environment based on the global map and the vehicle-centric local view, and demonstrate that the two parts can improve each other. The proposal does not require tedious and time-consuming environmental sampling and searching, but directly predicts the feasible path through end-to-end network. To be specific, global and local view are first performed by contractive autoencoder to learn latent obstacle avoidance knowledge. Then, global planner provides global information guidance for the local planners and ensemble planner composed of the two predicts feasible paths through an iterative update mechanism. Compared with general learning-based methods, our ensemble strategy shows stronger robustness in the test environments. Finally, we conduct extensive experiments on synthetic and real-world datasets, and the results demonstrate a competitive performance of the proposed method to classic planners and the latest ones.