Abstract: With the rapid development of intelligent agriculture, the apple industry in China has achieved large-scale production. However, the picking link still relies heavily on manual labor, facing challenges such as high labor costs, labor shortages, and difficulty in mechanized operation in hilly and mountainous areas. As the core component of apple picking robots that directly interacts with fruits, the end-effector’s flexibility and control accuracy are crucial to improving picking efficiency and reducing fruit damage. To address the problems of insufficient active flexibility, poor adaptability to different apple sizes, and weak ability to resist complex disturbances in traditional end-effectors, this study proposes a comprehensive technical solution combining structural design optimization and advanced control strategy. Firstly, a lightweight and adjustable flexible end-effector is designed. Adopting a three-jaw structure to simulate human grasping movements, the end-effector uses a micro servo electric cylinder as the driving mechanism, which converts the translational motion of the connecting rod into the opening and closing motion of the flexible gripper. The key innovation lies in the four-speed adjustable gear structure on the passive finger base, which realizes rapid gear switching within 15 s through a tenon-mortise slot connection, enabling the end-effector to safely grasp apples with diameters ranging from 20mm to 110mm. The core components are processed by 3D printing technology, and the gripper fingers are made of silica gel-fiber composite soft material with fin effect, effectively protecting the apple peel from damage. Secondly, an ADRC-MPC force-position hybrid control architecture integrating Position-Velocity Calculation (CPV) and Linear State Error Feedback (LSEF) is constructed to solve the problem of low control accuracy caused by friction disturbances in the transmission mechanism. In this architecture, the Extended State Observer (ESO) of ADRC is used to real-time estimate and compensate for total disturbances including friction, model errors, and external interference; the CPV module replaces the traditional Tracking Differentiator (TD) to avoid phase lag and noise sensitivity, establishing an accurate mapping relationship between force, displacement, and velocity through kinematic equations and Newton’s second law; the LSEF module serves as the coordination hub between MPC and ESO, dynamically correcting prediction deviations through speed error terms and introducing measured forces to enhance disturbance suppression capabilities; the MPC module optimizes the control strategy through a rolling time-domain optimization mechanism, embedding disturbance feedforward compensation items in the prediction model to achieve optimal control under constraints. Finally, simulation and physical experiments are conducted to verify the effectiveness of the proposed scheme. Simulation experiments based on Matlab show that the derived forward and inverse kinematic models of the end-effector are highly consistent with the simulation results, providing a reliable theoretical basis for practical control. Physical experiments are carried out under four gradient target grasping forces (2.5 N, 3.0 N, 3.5 N, 4.0 N) using standard apples with a diameter of 55.59 mm. The experimental results indicate that the system has excellent dynamic response and control accuracy: the average force tracking response time is 0.85 s, the average steady-state establishment time of contact force is 3.08 s, and the dynamic force control error is stably maintained within ±0.1N. Ablation experiments confirm that the MPC module significantly shortens the force tracking response time and convergence time, while the LSEF module ensures engineering practicality and parameter robustness with a simple linear structure. Comparative experiments with ADRC+PSO and DLADRC-OBLHOA control schemes show that the proposed ADRC-MPC architecture outperforms the comparison schemes in both steady-state arrival time and force tracking response time, with a reduction of 35.3% and 46.2% respectively compared to the ADRC+PSO scheme. This study not only breaks through the limitations of traditional end-effectors with fixed sizes but also improves the adaptability of the control strategy to complex disturbances in actual picking environments, providing a feasible and efficient technical solution for the flexible and precise picking of apple-harvesting robots, and also laying a foundation for the application of similar technologies in the harvesting of other ellipsoidal fruits such as pears and citrus.