Explainable Machine Learning for Autonomous Vehicle Positioning Using SHAP

Despite recent advancements in Autonomous Vehicle (AV) technology, safety remains a significant challenge for commercialization and development. The navigation system is crucial for AV safety, with road localization relying on accurate Global Navigation Satellite System (GNSS) positioning. However, GNSS may experience signal loss in urban environments. We previously proposed the Wheel Odometry Neural Network (WhONet) to provide continuous positioning in GNSS signal absence, integrating GNSS output with wheel encoders’ measurements. To ensure safety, qualitative assessment of WhONet predictions is needed, necessitating explanation for its reliability. We use Shapley Additive exPlanations (SHAP) to examine WhONet decision-making on an Inertial and Odometry Vehicle Navigation Benchmark Data subset. Our study highlights that rear wheels contribute most to position uncertainty error during approximate straight-line motion.