Aiming to address the dual challenges of declining localization accuracy and harvesting difficulties caused by fruit branch occlusion in complex agricultural environments, a reinforcement learning-based occlusion suppression and path planning method for apple picking robots was proposed. By constructing a vision-motion collaborative active perception system, a dynamic occlusion suppression calculation model (OSCM) at the visual perception layer was developed based on YOLO v8 model. This model integrated color histogram-based region segmentation and boundary fitting estimation algorithms to generate real-time occlusion rates and suppression weights heatmaps for target fruits. At the motion decision layer, a depth camera was rigidly mounted on the robotic arm's end-effector, establishing a three-dimensional occlusion suppression matrix mapping mechanism. This mechanism discretized the depth space from the camera to the target into a three-dimensional grid matrix, where each grid cell contained the OSCM-derived occlusion rate and suppression priority. Building on this framework, a deep reinforcement learning-driven obstacle avoidance decision method was proposed. It employed a composite reward function incorporating occlusion rate penalties and harvesting timeliness, enabling the double deep Q-network (DDQN) to search for the optimal suppression path within the three-dimensional grid. Experimental results demonstrated that in high-occlusion scenarios with branch-leaf density no less than 60% , the system reduced the average occlusion rate by 33% , increased the harvesting success rate from 67% to 94.7% , and decreased the average picking time per fruit by 3.2 s. The research pioneered the closed-loop integration of occlusion suppression mechanisms with reinforcement learning decision-making, providing a novel methodology for enhancing agricultural robots'adaptability in dynamic environments.