Versano and Klein developed a hybrid navigation filter for unmanned ground vehicles that trains a neural network exclusively on simulated data to predict measurement noise characteristics, then feeds those predictions into a classical unscented Kalman filter. The approach bypasses the standard bottleneck of deep-learning navigation — the need for extensive real-world datasets with perfect ground truth — by generating unlimited labeled examples in simulation. Testing across three datasets spanning different vehicles, sensors, road surfaces, and environments yielded a 12.7% position improvement over adaptive model-based approaches.
The structural insight is that the neural network is not replacing the filter — it is informing the filter about conditions the filter cannot sense. The Kalman filter already contains the correct physics for state estimation; what it lacks is real-time knowledge of how noisy its own inputs are. The network observes patterns in raw sensor data that correlate with noise regimes — vibration signatures, signal characteristics, dynamic conditions — and translates those observations into the noise parameters the filter needs. The learning operates on the metadata of the measurement, not on the measurement itself.
This division of labor — classical algorithm for the physics, learned component for the conditions — resolves a tension that appears across engineered systems. Climate models use physics for dynamics and machine learning for subgrid parameterization. Trading systems use financial theory for pricing and statistical models for volatility estimation. The pattern is: use the principled model for what it knows and the learned model for what varies, and connect them at the interface where knowledge of conditions is needed.
(arXiv:2603.11649)