Bridging the Sim-to-Real Gap in a Small-Scale Autonomous Platform: Experimental Assessment of Localization Robustness on an Autonomous Go-Kart Platform

Abstract

Autonomous systems often perform reliably in controlled environments but show degraded performance when deployed in real-world conditions. This difference is commonly referred to as the sim-to-real gap. In this thesis, the term is used in a system-level sense, since the AP4 autonomous go-kart platform depends on vehicle models, sensor calibration, localization assumptions, and controlled test conditions whose validity may change during field operation. The aim of the thesis is to investigate and bridge this gap, with focus on localization robustness, sensor fusion, and field deployment. The work evaluates and improves the AP4 platform’s existing localization pipeline. The main components considered are steering geometry, wheel-encoder odometry, inertial measurement unit (IMU) integration, and sensor fusion using an Extended Kalman Filter (EKF). Steering calibration and kinematic bicycle-model adjustments were used to improve the odometry estimate, while EKF parameter tuning was performed to improve sensor fusion performance and localization consistency. Localization performance was evaluated using motion-capture (MOCAP) measurements as ground truth and field tests at a go-kart track to assess behavior under more realistic operating conditions. The results show that the odometry and EKF-based localization were improved through steering calibration, kinematic model adjustments, and EKF tuning. Compared to odometry, the EKF-filtered estimate provided more accurate heading, improved trajectory consistency, and more stable long-term localization. However, reliable autonomous driving was still limited by higher-level functions outside the main scope of this thesis, particularly LiDAR-based SLAM, Nav2 behavior, and trajectory control. This indicates that reducing the sim-to-real gap of the AP4 platform requires continued validation of the complete localization and navigation stack under realistic field conditions.