Behavior-based search and surveillance algorithm for aerial robotic swarms

Resumen

Multi-robot systems have clear advantages over their single-agent counterpart, given that they are able to parallelize the assigned task to speed it up, and in case of failures or loss of agents, the mission can still be accomplished thanks to their redundancy characteristics. However, they also present some inconveniences that must be considered, such as the difficulties to coordinate the agents so that the entire group works together as a team. With the reduction of hardware costs and the increase of communication and computation capabilities, it has been possible in the last decade to solve some tasks with increasing number of robots. Nowadays, some lines of investigation are devoted to develop algorithms to control even hundreds of them, which are usually called robotic swarms and have their own specific characteristics. These advancements have also reached the aerial versions of these systems, which can be used for different missions, for instance, the search and the surveillance of areas in open spaces. This dissertation is devoted to present a novel algorithm to address search and surveillance missions of flat-and-rectangular-like areas with aerial robotic swarms (more specifically, quadcopters) equipped with downward cameras. The algorithm is based on a behavioral network made up by 6 different modules that act together to decide where to move during the mission. These modules, or behaviors, use the data sensed by the agent itself, and the information received by broadcast messages sent using radio frequency modules from the other agents. The algorithm proves to be scalable to tens of agents, robust against different type of failures and errors, and it is flexible to adapt to different scenarios. The algorithm has the drawback that its configuration is not a straightforward process, since optimal configurations must be found for a large variety of scenarios. This is solved by combining an optimization based on a genetic algorithm applied in a limited number of specific scenarios, and a Gaussian Process to model the optimal values as a function of the scenario characteristics. This algorithm has been tested in both simulations and with real robots, in an indoor arena and outdoors. The majority of the objectives established at the beginning of this work is satisfactorily fulfilled. Further investigations are also proposed at the end of this document to enlarge the capabilities of this approach.