Learning-Enabled Output–Feedback Design for Lane Change Control: A Comparative Study

Autonomous lane changing in real-world driving is challenging because sensor noise and practical sensing constraints often limit the availability of reliable full-state measurements. Although many model-based and learning-based control approaches assume access to full-state information, such assumptions may not hold in realistic autonomous driving environments. This paper presents an experimental study comparing control strategies that use a learning-enabled estimator-based output-feedback approach for autonomous lane change maneuvers under realistic state measurement conditions. A small-scale remotely controlled (RC) car platform is used as a scaled experimental testbed to evaluate a model-based model predictive control (MPC) controller, a learning-based state feedback controller, and an estimator-based output-feedback learning controller using real sensor data. The estimator-based controller learns an optimal steering policy online without direct access to full system states. Experimental results show that the estimator-based output feedback demonstrates improved robustness to sensor inaccuracies and external disturbances, smoother lane transitions, and more reliable adaptation in safety-critical scenarios when state measurements are limited or noisy. These results provide experimental evidence that learning-enabled estimator-based output feedback can reduce reliance on expensive sensing configurations while maintaining reliable control performance, supporting its potential practical relevance for autonomous driving systems operating under limited measurement conditions.