Abstract:
In this paper, we present a framework for collaborative uncertainty-aware navigation for swarms of vision based multirotor micro aerial vehicles (MAV). We assume that each MAV in a swarm is equipped with a forward-facing monocular camera, and that the vehicles are capable of using feature data to map the environment and perform vision based localization. Additionally, the vehicles are also capable of computing relative poses between each other in order to improve accuracy of pose estimation. For this scenario, we develop a navigation framework which seeks to improve the reconstructed maps and plan trajectories such that localization uncertainty is minimized. Within this framework, we first utilize an evolutionary algorithm that generates better viewpoints for the MAVs from which the map of the environment can be improved. This generated map is subsequently used as a source of information to perform path planning for each vehicle using a rapidly exploring random belief tree. This algorithm, while connecting start and goal poses with collision-free trajectories, ensures that the vehicles prioritize observing feature-rich areas and never lose sight of features, thus improving localization accuracy. Additionally, the algorithm is also capable of estimating when corrections would be required through relative poses and where these observations should be obtained, such that one vehicle can improve the accuracy of its neighbors. Through these approaches, we generate smooth, uncertainty-aware paths that are suitable for MAV navigation.