Safety-Oriented Comparative Assessment of Autonomous Path-Tracking Controllers Using a Real-World-Validated Virtual Vehicle Model

Comparisons of path-tracking controllers based mainly on tracking error may overlook differences in steering activity, vehicle response, and lane containment. This study compares Pure Pursuit, Stanley, and Model Predictive Control using a common virtual vehicle model and multiple safety-related performance indicators. The virtual vehicle model was evaluated against two separate CAN recordings from the same vehicle without recalibration between evaluations. Vehicle-speed and front-wheel-speed predictions showed R2 values above 0.97 in both recordings; yaw-rate R2 values were 0.957 and 0.930, whereas lateral-acceleration agreement was lower, particularly in the second recording. A Common Path Manager provided consistent reference-path information and steering constraints for all three controllers. They were evaluated at 5 and 10 km/h on idealized hairpin and spiral paths using tracking errors, steering-command-rate RMS, kinematic-response measures, and center-of-gravity-based and sampled vehicle-body lane-containment measures. MPC achieved the lowest cross-track-error RMSE in both scenarios, whereas Pure Pursuit produced the lowest steering-command-rate, lateral-acceleration, and yaw-rate RMS values. Stanley achieved the lowest heading-error RMSE in the spiral scenario. In the hairpin tests, the vehicle center of gravity remained within the lane for all controllers, although the body was partially outside the modeled lane corridor in 37.048–53.352% of evaluated samples. These findings show that lower tracking error does not necessarily coincide with lower steering-command activity or improved CG-based lane keeping. The evaluated quantities are safety-related performance indicators for the investigated simulation conditions and do not constitute a comprehensive assessment of vehicle safety.