A Predictive Deep Learning Approach to Output Regulation: The Case of Collaborative Pursuit Evasion

TL;DR

This paper introduces a deep learning-based control method for autonomous aerial vehicles in pursuit-evasion scenarios, enabling adaptive target tracking without prior knowledge of the evader's strategy.

Contribution

It presents a novel predictive control law using deep neural networks to approximate adversarial behavior in unknown environments for underactuated systems.

Findings

Effective pursuit evasion demonstrated in simulations

Deep neural network accurately predicts evader behavior online

Control law adapts to adversarial strategies in real-time

Abstract

In this paper, we consider the problem of controlling an underactuated system in unknown, and potentially adversarial environments. The emphasis will be on autonomous aerial vehicles, modelled by Dubins dynamics. The proposed control law is based on a variable integrator via online prediction for target tracking. To showcase the efficacy of our method, we analyze a pursuit evasion game between multiple autonomous agents. To obviate the need for perfect knowledge of the evader's future strategy, we use a deep neural network that is trained to approximate the behavior of the evader based on measurements gathered online during the pursuit.