Depth-based Reactive Quadrotor Motion Planning and Control in Diverse Environments

Abstract

This thesis explores motion planning and control strategies for enabling rapid quadrotor navigation in unknown environments using only limited field-of-view depth sensors. We propose real-time onboard algorithms that enable agile flight through diverse and cluttered spaces. First, we present a reactive planner based on forward-arc motion primitives that uses a short history of RGB-D observations to safely navigate near obstacles. A safe stopping strategy ensures that the quadrotor always maintains a trajectory that allows it to hover safely within known free space. Second, we build on recent advances in reinforcement learning with differentiable physics to develop a navigation policy that predicts thrusts directly from depth and state observations. We show that by using privileged information during training, our approach is able to navigate around large obstacles. Through extensive simulation and real-world experiments, we show that our methods outperform baselines in both speed and reliability in cluttered environments. These results contribute toward the deployment of agile autonomous flight in real-world applications such as search and rescue and exploration.