Guidance laws for partially-observable UAV interception based on linear covariance analysis

Abstract

Unmanned Aerial Vehicles (UAVs) have proliferated the skies in recent years as they have become extremely popular for all different kinds of commercial, government, and recreational usage. With all this activity, there remains an open security threat, particularly to airports, soldiers, and large crowds of people. This thesis work is motivated by the idea of an autonomous pursuer drone that can intercept and capture a malevolent drone. Due to the limited payload of drones, we consider pursuit-evasion games characterized by partial state observability. Specifically, we consider bearing-only measurements, as can easily be obtained from a single camera sensor. In this work, an optimal control formulation for a drone pursuit-evasion game is achieved in 7 states. Using the sophisticated Continuous Computation and Compression (C3) library, a new optimal controller is calculated in compressed tensor train (TT) format. By compressing the state space, it is possible to calculate the optimal control action at any state in real time. A set of observability maneuvers is identified to help the pursuer improve the estimate quality of an Unscented Kalman Filter (UKF) tracking the target's relative position and velocity. Using Linear Covariance Analysis, an novel algorithm is developed to pick the series of maneuvers that gives the best probability of capture. This algorithm is demonstrated on a quadrotor in flight intercepting a simulated evader drone, and it is shown to improve the tracking performance error by several orders of magnitude.