Architectural Comparison of Kalman Filter-Based Multi-Sensor Localization for a Duckiebot

Accurate localization is a fundamental requirement for autonomous mobile robots, as it directly impacts navigation performance and system reliability. In low-cost robotic platforms, localization must rely on onboard sensors such as wheel encoders, inertial measurement units, and vision-based systems, which are inherently affected by noise, drift, and limited observability. These challenges motivate the use of sensor fusion techniques to obtain reliable state estimates over time. This thesis investigates how different Kalman filter architectures influence localization performance for a unicycle-model mobile robot. Rather than proposing a novel estimation algorithm, this work focuses on the role of architectural design choices in sensor fusion. Standalone filtering baselines, centralized sensor fusion, cascaded filtering schemes, and error-state Kalman filter architectures are analyzed within a unified estimation framework, with particular attention to accuracy, robustness, and fault tolerance when fusing heterogeneous sensor measurements. The considered architectures are implemented in a MATLAB/Simulink simulation environment that enables consistent modeling of robot kinematics, sensor characteristics, and noise processes. Wheel encoder and inertial measurements provide high-rate motion information, while vision-based localization using AprilTag detections acts as an external correction source to mitigate long-term drift. The performance of each filtering approach is evaluated under realistic operating conditions using quantitative error and consistency metrics. Simulation results are further validated through experimental tests on a Duckietown robotic platform, demonstrating the practical relevance of architectural choices in Kalman filter-based localization for real-world mobile robots.