Abstract: To address the issue of low assembly efficiency caused by prolonged hole-searching stages in robot peg-in-hole assembly, a robot active posture deviation trajectory optimization method integrating random forests (RF) with hybrid Kepler optimization algorithm (KOA) and dung beetle optimizer (DBO)is proposed. Firstly, the peg-in-hole contact force data are collected through raster motion, and the dataset is constructed and divided into deflection domains. Secondly, the KOA is employed to optimize the RF to build a force-deflection domain classification model, while the DBO is used to optimize the RF for constructing a force-deflection regression model. Finally, based on the deflection domain predicted by the classification model, the robot is controlled to perform active posture deviation motion to accurately find the hole direction. The experimental results demonstrate that the proposed force-position vector model reduces the hole-searching time by 13.3 s and decreases the number of steps required by 50% compared to the pre-optimized version, effectively enhancing the hole-searching efficiency while optimizing the trajectory of the search process.