This dissertation addresses the problem of non-myopic online exploration and visual sensor coverage of large-scale unknown environments using an autonomous robot. We introduce a novel perception roadmap, referred to as the Active Perception Network (APN), that represents a connected configuration space over a concurrently built spatial map. The APN is modeled by a hierarchical topological hypergraph that equips a robot with an understanding of how to traverse throughout a concurrently built spatial map, and facilitates predictive reasoning on the expected visible information of the environment from untraversed regions of the map.
As new information is added to the map during exploration, the APN is iteratively updated by an adaptive algorithm entitled Differential Regulation (DFR), which applies difference-aware strategies to constrain the complexity of each update to the size of changed map information, independent of its total size. DFR employs a view sampling-based strategy to expand and refine traversability knowledge as map knowledge increases, using a novel frontier-based approach to evaluate information gain and guide the sampling and pruning of views within the APN. The APN serves as a knowledge model which can be applied for graph-based exploration planning. An evolutionary planner, designated as APN-P, leverages the hierarchical representation of the APN to perform non-myopic exploration planning that dynamically adapts to the changing map and APN states.
This dissertation further presents a software development framework, Active Perception for Exploration, Mapping, and Planning (APEXMAP), that addresses the unique software design and implementation challenges inherent to online exploration and active perception tasks, which are non-trivial. APEXMAP provides a generalized modular framework for these challenges, which is made open source for the benefit of the research community.