Environment-Aware Learning of Smooth GNSS Covariance Dynamics for Autonomous Racing

A team from Caltech has introduced LACE (Learning-based Adaptive Covariance Estimation), a novel AI framework that dynamically models GNSS (Global Navigation Satellite System) measurement uncertainty for autonomous racing vehicles. Presented in a paper accepted to ICRA 2026, the system addresses a critical challenge in high-speed autonomy: providing accurate, stable, and smooth state estimation when GPS signals become unreliable due to environmental factors like urban canyons or tunnels. Unlike traditional static models, LACE treats covariance evolution as an exponentially stable dynamical system, allowing it to adapt in real-time to changing conditions.

The technical core of LACE uses a deep neural network (DNN) with an attention mechanism to predict process noise from environmental features. By systematically imposing spectral constraints based on contraction theory, the researchers provide formal mathematical guarantees of exponential stability and temporal smoothness for the resulting covariance dynamics. This means the estimated uncertainty won't jump erratically, which is vital for stable vehicle control. The framework was validated on a real AV-24 autonomous racecar, demonstrating superior localization performance and smoother covariance estimates in GNSS-degraded scenarios. This work bridges the gap between perceived uncertainty in sensing and the sensitivity of control algorithms, paving the way for safer, more reliable autonomous systems operating at the limits of performance.