SONID: underwater sonar SLAM integrated with neural inertial-DVL odometry

We present SONID, an underwater SLAM framework that tightly integrates neural inertial-DVL odometry with forward-looking sonar perception. SONID employs a lightweight convolutional neural network to adaptively estimate noise covariances for an Invariant Extended Kalman Filter, enabling robust and accurate state estimation in dynamic underwater environments. Multi-sensor constraints from odometry, sonar features, and loop closures are formulated in a factor graph and optimized incrementally using ISAM2 for globally consistent trajectory and mapping. Simulation results show that SONID achieves high localization precision, strong tracking success rates, and reliable mapping, demonstrating its effectiveness for challenging underwater SLAM tasks.