This thesis focuses on the combination of the observation problem with the patrol problem within a multi-agent framework. The observation problem aims to track targets by maximizing the number of mobile targets observed by at least one agent over time. While the patrol problem aims to visit a set of locations in the environment as frequently as possible, essentially achieving regular coverage of the environment.
We propose that the agents face an exploration-exploitation dilemma, achieved through actively searching for new targets via environmental patrols and maximizing target observation. We term this novel challenge the Patrolled Observation Problem (POP). By not solely concentrating on target tracking, the agents also mitigate the risk of manipulation by intelligent targets that might attempt to influence their movements.
We present a set of solutions for the POP, relying on either potential fields (I-CMOMMT) or learning approaches, whether through reinforcement (FFRL, F2MARL) or supervised learning (MALOS). These methods are compared with other strategies from existing literature. The experiments are initially conducted within a Gazebo/ROS2 simulation environment, where agents are represented by drones and targets by mobile ground robots. Subsequently, these methods are implemented and evaluated on actual drones within an aviary.
The thesis introduces two additional contributions in the form of tools, complementing the approaches to solving the POP. The first tool enables agents to efficiently identify locations with potential interest to be visited. The second tool aims to optimize mission parameters, such as the number of agents, while complying with one or more user-defined performance criteria.
In the interest of reproducibility, all codes, whether for the simulation environment or the methods themselves, are open-source.