Large-scale multi-robot 3D reconstruction of indoor environments poses challenges for robotics, virtual reality, construction, and more applications. This thesis proposes a scalable system architecture to advance the distributed multi-robot 3D reconstruction field and addresses problems in robot localization, coordination, and vision processing. A decentralized and communication-based multi-agent path finding framework is first developed to coordinate path planning among fleets of mobile robots through reinforcement learning. We further improve the communication overhead by a Field-of-View (FOV) study and selectively sharing information between neigh-boring agents. The FOV study reveals that smaller FOV sizes significantly reduce computational costs without compromising performance.

Add...[ Read more ]

Large-scale multi-robot 3D reconstruction of indoor environments poses challenges for robotics, virtual reality, construction, and more applications. This thesis proposes a scalable system architecture to advance the distributed multi-robot 3D reconstruction field and addresses problems in robot localization, coordination, and vision processing. A decentralized and communication-based multi-agent path finding framework is first developed to coordinate path planning among fleets of mobile robots through reinforcement learning. We further improve the communication overhead by a Field-of-View (FOV) study and selectively sharing information between neigh-boring agents. The FOV study reveals that smaller FOV sizes significantly reduce computational costs without compromising performance.

Additionally, an efficient 3D reconstruction pipeline is developed utilizing 360deg cameras. This leverages equirectangular projection conversion to facilitate deep learning-based 3D reconstruction models. Experimental evaluations demonstrate that 360deg camera data achieves more accurate and efficient scene reconstruction than perspective cameras. In summary, communication in multi-agent path finding, 360deg imaging processing, and vision-based 3D reconstruction have been explored and developed in this work. The resulting framework facilitates fleet-scale 3D reconstruction applications in site monitoring and beyond.