PRML2 hybrid ML-Kalman filter improves vehicle localization during sensor outages
New framework uses onboard sensors and physics-regularized learning for robust pose estimation.
Accurate vehicle localization is critical for autonomous driving, but systems relying on GPS or IMU often fail during signal outages or low-friction conditions. A new paper from TU Munich—accepted at IROS 2026—introduces PRML2 (Physics-Regularized Machine Learning for Localization), a hybrid framework that fuses classical Kalman filtering with data-driven learning. The key innovation is an end-to-end differentiable Kalman filter that trains an ML model to stay physically consistent with vehicle motion models, preventing the drift common in pure learning-based approaches.
PRML2 relies solely on cheap onboard sensors (IMU, wheel speed) already present in production vehicles, making it a cost-effective backup. On public datasets, it demonstrates superior accuracy over both pure Kalman filters and black-box ML methods, while remaining fast enough for real-time deployment. The authors also release a new dataset collected on low-friction surfaces (e.g., ice) to challenge and benchmark upcoming localization algorithms. The work shows that blending physics priors with learned corrections can bridge the reliability gap between classical and neural approaches for autonomous navigation.
- PRML2 uses a differentiable Kalman filter to train an ML model end-to-end with physics constraints, improving consistency with vehicle motion models.
- Outperforms pure Kalman filters and standalone ML methods on public datasets, and runs in real-time on standard hardware.
- Includes a novel low-friction dataset (ice, snow) to support vehicle localization research under degraded conditions.
Why It Matters
A robust, low-cost localization backup for autonomous vehicles when GPS fails, enhancing safety and reliability.