Belief-space Planning for Active Visual SLAM in Underwater Environments

Autonomous mobile robots operating in a priori unknown environments must be able to integrate path planning with simultaneous localization and mapping (SLAM) in order to perform tasks like exploration, search and rescue, inspection, reconnaissance, target-tracking, and others. This level of autonomy is especially difficult in underwater environments, where GPS is unavailable, communication is limited, and environment features may be sparselydistributed. In these situations, the path taken by the robot can drastically affect the performance of SLAM, so the robot must plan and act intelligently and efficiently to ensure successful task completion. This document proposes novel research in belief-space planning for active visual SLAM in underwater environments. Our motivating application is ship hull inspection with an autonomous underwater robot. We design a Gaussian belief-space planning formulation that accounts for the randomness of the loop-closure measurements in visual SLAM and serves as the mathematical foundation for the research in this thesis. Combining this planning formulation with sampling-based techniques, we efficiently search for loop-closure actions throughout the environment and present a two-step approach for selecting revisit actions that results in an opportunistic active SLAM framework. The proposed active SLAM method is tested in hybrid simulations and real-world field trials of an underwater robot performing inspections of a physical modeling basin and a U.S. Coast Guard cutter. To reduce computational load, we present research into efficient planning by compressing the representation and examining the structure of the underlying SLAM system. We propose the use of graph sparsification methods online to reduce complexity by planning with an approximate distribution that represents the original, full pose graph. We also propose the use of the Bayes tree data structure—first introduced for fast inference in SLAM—to perform efficient incremental updates when evaluating candidate plans that are similar. As a final contribution, we design risk-averse objective functions that account for the randomness within our planning formulation. We show that this aversion to uncertainty in the posterior belief leads to desirable and intuitive behavior within active SLAM.