Visual Localization and Policy Learning for Robotic Large-Diameter Peg-in-Hole Assembly Tasks

The conventional component assembly techniques employed in manufacturing industries typically necessitate laborious manual parameter calibration prior to system deployment, while existing vision-based control algorithms suffer from limited adaptability and inefficient learning capabilities. This paper presents a novel framework for automated large-diameter peg-in-hole assembly through convolutional network-based perception and reinforcement learning-driven control. Our methodology introduces three key innovations: (1) an enhanced deep segmentation architecture for precise identification and spatial localization of peg-end centroids, enabling accurate preliminary peg-in-hole; (2) a hybrid control strategy combining deep deterministic policy gradient (DDPG) reinforcement learning with classical control theory, augmented by real-time force feedback data acquisition; (3) systematic integration of visual–spatial information and haptic feedback for robust error compensation. Experimental validation on an industrial robotic platform demonstrates the method’s superior performance, achieving an Intersection over Union (IoU) score of 0.946 in peg segmentation tasks and maintaining insertion stability with maximum radial forces below 5.34 N during assembly operations. The proposed approach significantly reduces manual intervention requirements while exhibiting remarkable tolerance to positional deviations (±2.5 mm) and angular misalignments (±3°) commonly encountered in industrial assembly scenarios.