Kinematics-Aware Next-Best-View Planning for Autonomous Underwater Vehicles
My thesis: coupling Bayesian search with kinematics-aware next-best-view inspection so an underwater vehicle can find and inspect seabed targets without an exhaustive coverage survey.
- thesis
- robotics
- planning
- underwater
- ros2
Read the full thesis · DOI: 10.26153/tsw/66342
Abstract
M.S.E. thesis, The University of Texas at Austin, 2026. Supervisor: Mitch Pryor.
Autonomous underwater vehicles (AUVs) inspecting multiple seabed targets must both find them and plan informative views of them under severe sensing and kinematic constraints. Conventional coverage surveys meet this by exhaustively sweeping the seabed, which guarantees completeness but scales poorly and spends most of a mission imaging empty ground; next-best-view (NBV) planning promises to inspect adaptively, but has largely been studied apart from the search that must precede it and the kinematic limits an AUV imposes.
This thesis develops a reduced-order simulation coupling multi-target Bayesian search with kinematically-aware NBV inspection for an underactuated, non-holonomic AUV carrying a forward-looking sonar. It asks whether the resulting adaptive mission earns a net benefit over exhaustive coverage once these constraints hold. The work makes four contributions. The first is a modular ROS 2 simulation workspace whose vehicle, mission, and environment packages are all parameterizable. The second is a behavior-tree mission architecture, built on BehaviorTree.CPP, in which a detected target interrupts the search on its own, with no hand-written transition logic. The third extends the NRG lab’s NBV planner with a principled stopping rule and a utility that trades expected information gain against the true kinematic cost of reaching each view, weighted by a single tunable parameter. The fourth is the cumulative information ratio (CIR), a metric that rewards seeing each target’s faces from many angles with diminishing return and scores the adaptive framework and the coverage baseline on equal terms.
Evaluated over two viewpose sampling manifolds, a cone and a helix, across a scene of clustered and isolated targets, the adaptive mission proved substantially faster to complete and quicker to saturate its information gain, conceding only a modest, bounded deficit in completeness. The manifold, more than the weighting, set the size and reliability of that trade, and an occlusion reversal in the per-target results showed completeness under the greedy, gain-terminated policy to depend on a target’s local relief rather than the terrain occlusion it was built to test. Reconstruction quality proved decoupled from the information score: coarse at operational standoff and set chiefly by the manifold. Read as a deterministic upper bound, these results establish calibrated expectations for the probabilistic sensing, richer terrain, and field deployment that this modular framework was built to support.