Human-Centered Corrective Control in Semi-Automatic Construction Machinery: Enhancing Stability and Operator Intent Integration

Semi-automatic piloting combines prerecorded machine motion with operator corrections, but how corrective inputs evolve across repeated interactions remains unclear. We examined corrective lever inputs during simulated grading and loading with two skilled excavator operators. Each operator corrected their own and the other operator’s recorded actions, with corrections either immediately applied (Apply) or recorded without affecting the current motion (No Apply). Cycle-level root mean square (RMS) values summarized corrective-input magnitude, and generalized estimating equations examined temporal trends. RMS generally decreased during grading under Apply but tended to increase during loading, particularly along the Swing axis. The grading model detected feedback-dependent slopes, whereas neither task model detected baseline-source-dependent slopes. These findings do not establish equivalence between baseline sources, and the task comparison remains descriptive. Because the experiment involved two operators, fixed task order, and a simulator, the results are exploratory. Estimates remain conditional on the assumed independence between repeated series from the same operator. RMS quantifies corrective-input magnitude and provides complementary behavioral information for assessing task performance; the prominent Swing-axis corrections observed during loading identify a concrete target for evaluating assistance in semi-automatic excavators.