AI Improves Underwater Navigation with Geometry and Simulation

Autonomous vehicles, from underwater drones to self-driving cars, rely on precise positioning to navigate complex environments. Inertial sensors, which measure motion, are crucial for this task, but their accuracy often degrades due to noise and real-world uncertainties. Traditional s for filtering this noise, like the Kalman filter, have been enhanced with adaptive techniques, but these approaches struggle with the geometric complexities of motion in three-dimensional space. A new study addresses this by integrating neural networks with a geometric filtering framework, demonstrating significant improvements in navigation accuracy for autonomous underwater vehicles (AUVs), a challenging domain where precise positioning is essential for tasks like ocean exploration and infrastructure inspection.

The researchers developed a neural-aided adaptive invariant Kalman filter (IKF) that outperforms existing s in position accuracy. In tests on real-world underwater navigation data, their approach reduced the root mean square error (RMSE) in position by 17.1% compared to a standard adaptive extended Kalman filter (EKF). This improvement stems from combining two key innovations: a geometric formulation that respects the natural structure of motion on Lie groups, and a lightweight neural network trained entirely in simulation to estimate noise parameters from raw sensor data. was evaluated using the A-KIT dataset, comprising 80 minutes of real AUV missions, where it consistently achieved lower errors across diverse trajectories, with mean RMSE values as low as 6.3 meters in some configurations.

To achieve this, the team derived a novel theoretical extension of classical innovation-based process noise adaptation, formulated directly within the Lie-group framework used by invariant Kalman filters. Unlike traditional Euclidean s, this approach preserves the geometric structure of motion, leading to more consistent error dynamics. They then designed a neural network with a simple architecture: three one-dimensional convolutional layers processing windows of 100 inertial measurements, followed by fully connected layers that output six noise parameters. This network, trained in a simulation-to-real (sim2real) framework, learns to estimate process noise covariance from simulated data with perfect ground truth, eliminating the need for labeled real-world data. The hybrid framework blends this neural estimate with an innovation-based adaptive term, using a weighting parameter set to 0.6 to balance data-driven and model-based contributions.

, Detailed in Table II of the paper, show that the proposed neural-aided adaptive IKF variants consistently outperform baselines. For example, on trajectory 2, achieved an RMSE of 2.9 meters compared to 3.8 meters for the adaptive EKF. The neural network, trained on 48 minutes of simulated data with diverse motion profiles and noise levels, generalized effectively to the real-world A-KIT dataset, indicating successful sim2real transfer. The paper notes that geometric invariance enhances learning-based adaptation, as the invariant framework provides a structurally superior foundation for noise modeling. This is evidenced by the fact that neural enhancement yielded only marginal gains in classical EKF setups but substantial improvements when integrated with the invariant filter, highlighting the importance of geometric consistency.

This advancement has practical for autonomous navigation in environments where GPS is unavailable, such as underwater or in dense urban areas. By improving positioning accuracy without requiring extensive real-world data collection, could lower costs and increase reliability for applications like marine research, underwater robotics, and autonomous logistics. The sim2real training approach is particularly valuable in domains like underwater navigation, where collecting labeled data is challenging and expensive. However, the study acknowledges limitations: the effectiveness depends on the representativeness of the simulated training domain, and the neural network adds computational complexity compared to fixed-noise implementations. Future work may explore broader applications and further optimization to address these constraints while maintaining the geometric and adaptive benefits demonstrated in this research.