Path tracking failure analysis of an autoware teaching vehicle in real world environments

This paper presents a systematic experimental analysis of path tracking failures in an Autoware-based teaching vehicle during turning maneuvers, using RTK-GPS trajectory data from 11 real-world runs. Three categories of failure sources are identified and quantified. Procedural sources (manual initialization inaccuracy) produce a straight-line drift of 5.685 m and a turn entry lateral error of 1.693 m. Environmental sources reveal that left-turn and right-turn curves have nearly identical radii ( \(\approx \) 11 m), but the narrower right-turn exit (4 m vs. 6 m) reduces lateral clearance to 1.1 m. Algorithmic sources show that the Pure Pursuit controller’s fixed lookahead distance (1.5 m) mismatches natural speed variation (7–9 km/h on straights, <3 km/h in curves), causing turn entry lag and 96% error growth. When error exceeds the lookahead distance, the controller enters positive feedback, producing a commanded curvature of 18.116 m \(^{-1}\) —61 times the vehicle’s physical limit (0.297 m \(^{-1}\) )—and the vehicle decelerates from 12.6 km/h to 0. The main contribution is a quantitative demonstration of a deterministic cascading failure chain: localization deviation, lateral error accumulation, lookahead violation, curvature explosion, steering saturation, and velocity collapse. This reveals that Autoware’s failure is not a single-source fault but a system-level coupling failure with positive feedback.