Multi-robot active information gathering using random finite sets

Many tasks in the modern world involve collecting information, such as infrastructure inspection, security and surveillance, environmental monitoring, and search and rescue. All of these tasks involve searching an environment to detect, localize, and track objects of interest, such as damage to roadways, suspicious packages, plant species, or victims of a natural disaster. In any of these tasks the number of objects of interest is often not known at the onset of exploration. Teams of robots can automate these often dull, dirty, or dangerous tasks to decrease costs and improve speed and safety. This dissertation addresses the problem of automating data collection processes, so that a team of mobile sensor platforms is able to explore an environment to determine the number of objects of interest and their locations. In real-world scenarios, robots may fail to detect objects within the field of view, receive false positive measurements to clutter objects, and be unable to disambiguate true objects. This makes data association, i.e., matching individual measurements to targets, difficult. To account for this, we utilize filtering algorithms based on random finite sets to simultaneously estimate the number of objects and their locations within the environment without the need to explicitly consider data association. Using the resulting estimates they receive, robots choose actions that maximize the mutual information between the set of targets and the binary events of receiving no detections. This effectively hedges against uninformative actions and leads to a closed form equation to compute mutual information, allowing the robot team to plan over a long time horizon. The robots either communicate with a central agent, which performs the estimation and control computations, or act in a decentralized manner. Our extensive hardware and simulated experiments validate the unified estimation and control framework, using robots with a wide variety of mobility and sensing capabilities to showcase the broad applicability of the framework