The research presents a multi-sensor underwater navigation framework that integrates physically consistent robot dynamics with visual–inertial odometry. A Port-Hamiltonian Neural ODE network learns energy-conserving underwater dynamics from thruster commands and pre-integrates them as factors in a tightly coupled factor graph, alongside visual and inertial cues. The system jointly estimates trajectory and external forces, explicitly separating commanded motion from disturbances such as currents or contacts. On real underwater datasets, the framework achieves consistent accuracy improvements over the baseline with minimal computational overhead. In simulation with injected forces, it reduces average pose error by over 60% and achieves near-perfect force estimation.

Маалла Я. (науч. рук. Колюбин С.А.) Improving underwater state estimation by fusing vio with neural ode-predicted dynamics // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2026]. URL: https://kmu.itmo.ru/digests/article/16618