PHIONet: Port Hamiltonian Inertial Odometry Network

Learning-based inertial odometry has shown strong potential for mitigating drift by extracting data-driven motion priors. However, existing methods primarily learn statistical mappings from inertial measurements to motion states, failing to consider the power balance between the specific-force input port and the predicted velocity. This limitation is particularly pronounced in aggressive MAV flight, where rapid accelerations, high angular rates, and aerodynamic dissipation can cause locally accurate velocity estimates to accumulate into trajectory drift. We propose PHIONet (Port Hamiltonian Inertial Odometry Network), which incorporates dissipative energy dynamics as a physically grounded prior. PHIONet translates the continuous port-Hamiltonian dynamics into a discrete energy-consistency residual over finite IMU windows, employing the DCM (Dissipative Coupling Module) to bridge IMU correction and velocity prediction. Experiments on the public Blackbird and EuRoC MAV datasets show that PHIONet achieves lower ATE and RTE than representative learning-based inertial odometry baselines. Code and related materials are available at https://anonymous.4open.science/r/PHIONet-D3CC.