Multi-robot adversarial patrolling: Handling sequential attacks

Robot teams are commonly used for security tasks, where they are required to repeatedly monitor an area in order to prevent penetrations, initiated by an adversary. Current research in this field focuses mainly on detecting penetration attempts, but not on responding. Requiring the robots to also inspect and handle the penetrations has a significant impact on the patrol, as each penetration attempt also influences the robots' behavior, making them vulnerable to multiple attacks. Moreover, a knowledgeable adversary can initiate two sequential attacks, where the second attempt exploits the vulnerable points caused by the requirement that a robot handle the first penetration attempt. In this work, we consider the problem of sequential attacks and examine different robot policies against such adversarial behavior. We provide an optimal patrol strategy for various penetration attempt patterns. Our novel approach considers a full history-length policy, while previous work only handled very limited lengths of history. The use of a longer history improves the results. Moreover, we show how to significantly reduce, in practice, the exponential space state of the problem, while maintaining the optimality of the solution.