In complex orchard environments, robotic fruit picking is challenged by branch occlusions, positioning inaccuracies, and low operational efficiency. A deep reinforcement learning-based framework for collision-free and efficient citrus harvesting with a robotic manipulator was proposed. The system integrated a YOLO 11 perception model for fruit classification and localization, together with a pressure prediction module for a flexible gripper to enable stable and damage-free grasping. An improved path planning algorithm, termed Attention LSTM HER SAC (ALH-SAC), was developed to achieve adaptive and reliable grasp control. The proposed algorithm extended the soft actor-critic framework by incorporating hindsight experience replay, an LSTM encoder for modeling temporal joint dependencies, and an attention mechanism to emphasize critical spatial features under occlusion. A customized reward function was designed to balance rapid target approach, obstacle avoidance, and optimal grasp orientation. In addition, a hybrid guidance strategy combining heuristic priors with a Cartesian-space linear interpolation prior was introduced to accelerate training convergence and enhance robustness. Simulation results demonstrated that ALH-SAC achieved a positioning accuracy of (10.35±4.18) mm, outperforming HER TD3 and HER DDPG baselines by over 12.5 mm on average. Orchard experiments further confirmed the effectiveness of the proposed method, achieving an average picking success rate of 84% and an average operation time of 10.58 s, validating its practical applicability in real-world citrus harvesting.