ON THE LIMITATIONS OF AGGREGATE TRACKING-ERROR METRICS FOR EVALUATING WEATHER ROBUSTNESS IN AUTONOMOUS ROW-FOLLOWING ROBOTS

Vision-guided agricultural robots must continue following crop rows as visual conditions degrade, yet robustness is often summarized by a single tracking-error statistic. This paper demonstrates, in simulation, that such a statistic can rank conditions incorrectly. A differential-drive robot with a fixedthreshold HSV row detector was simulated in Webots for 20 s of simulated time under clear weather (NORMAL), synthetic fog, and synthetic rain. Each steering command was also sent to a physical Arduino Uno, and an identifier-matched acknowledgement was used to measure the communication path. Fog reduced the mean number of detections per step from 4.5 to 0.6 and increased the RMS pixel error from 24.1 px to 46.9 px. Rain produced the lowest and most stable pixel error of the three conditions (16.1 px mean), yet the detector reported exactly five detections in all 624 steps, and a majority of detections had the same fixed width produced by the detector’s oversized-region split. These diagnostics indicate that the global brightness reduction in the rain model moved the detector’s effective dark-pixel threshold so that the mask merged into one static region. Thus, the low error reflects detector lock-on rather than accurate row following. We recommend reporting detection-stability diagnostics alongside tracking error when evaluating fixed-threshold perception pipelines under weather perturbations.