UC San Diego

Many real-world mobile robot applications, such as disaster response, military reconnaissance, and environmental monitoring, require operating in unknown and unstructured environments. This calls for algorithms that empower robots with active information gathering capabilities in order to autonomously and incrementally build a model of an environment. In this dissertation, we present a novel 3-D multi-class online mapping approach using a stream of range and semantic segmentation observations. Moreover, we derive a closed-form expression for the Semantic Shannon Mutual Information (SSMI) between our proposed map representation and a sequence of future sensor observations. Using an octree data structure, we reduce the memory footprint of the map storage for large-scale environments, while simultaneously accelerating the computation of mutual information. This allows real-time integration of new sensor measurements into the map, and rapid evaluation of candidate future sensor poses for exploration. Additionally, we introduce a differentiable approximation of the Shannon mutual information between grid maps and ray-based measurements, enabling gradient-based occlusion and collision-aware active mapping. The gradient-based active mapping in the continuous space of sensor poses reduces the optimization complexity from exponential in the number of robots to linear, paving the way for extension from a single agent to a team of robots. We formulate multi-robot exploration as a combination of multi-robot mapping and multi-robot planning, where both sub-problems are specified as an instance of multi-agent Riemannian optimization. We propose a general distributed Riemannian optimization algorithm that solves both mapping and planning in fully decentralized manner. Our method, named Riemannian Optimization for Active Mapping (ROAM), enables distributed collaborative multi-robot exploration, with only point-to-point communication and no central estimation and control unit. Lastly, we deploy our active mapping method on a team of ground wheeled robots in both simulation and real-world environments, and compare its performance with other autonomous exploration approaches.