Semantic-Aware Trajectory Planning for UAV in Dynamic Environments

Abstract

Generating trajectories that leverage semantic information to guide a UAV safely and accurately to its destination in a dynamic environment remains an open problem. In the existing literature, semantics have been used to prioritize certain areas - either to guide the UAV through or to avoid them -
for specific objectives, such as reducing errors in visual-inertial SLAM (VI-SLAM). However, prior work typically assumes a static environment when performing collision checking, even in cluttered and dynamic settings.

We propose a two-stage workflow: The first stage performs semantic-aware pathfinding. The second stage optimizes the resulting path, incorporating kinematic constraints and performing collision checking that accounts for obstacle motion, while still ensuring real-time performance.

To the best of our knowledge, this is the first approach that generates UAV trajectories by simultaneously leveraging semantic information and accounting for cluttered, dynamic environments. A summary video is available at https://youtu.be/I5w6AP7HThU