LAUSR

This paper presents a path planning algorithm for efficiently generating low-cost trajectories that meet mission requirements specified in Linear Time Logic (LTL), where cost functions are defined throughout the configuration space. The main idea of the paper is to increase efficiency by adding learning-based extensions to sampling-based path planning algorithms. The proposed method includes two layers: a high-level layer that determines how to expand the search tree to satisfy logical specifications and a low-level layer that extends the search tree to search for low-cost trajectories. To efficiently find low-cost trajectories, it is learned how to extend the search tree from the data via deep learning. By leveraging the conditional variational autoencoder, we learn the ideal search tree extension distribution in a given situation, which increases solution search efficiency. Simulations show that the proposed method finds a low-cost trajectory while meeting a given mission specification. Furthermore, it is confirmed that the proposed approach performs better than the existing methods.