Abstract: Objectives To address the challenges of traditional "sail-rudder" separated control in unmanned sailboat waypoint tracking tasks, such as mutual interference between control channels and strong conservative control, this paper proposes a deep learning-based "sail-rudder" joint model predictive control method for unmanned sailboats. Methods Firstly, the mathematical model of the unmanned sailboat's motion is established and the force conditions on the sail are analyzed. Then, a prediction model is constructed using a nonlinear state-space discretization method. The prediction model is identified online through a deep neural network, and multi-step prediction and output feedback correction techniques are employed to improve state prediction accuracy. Next, a composite objective function is formulated, integrating tracking error metrics and vessel speed metrics. Using cross-entropy optimization algorithms, the optimal control quantities for sail angle and rudder angle are solved within the prediction horizon, effectively overcoming the limitations of separated controller design. Results Finally, experiments were conducted on a PyTorch deep learning simulation platform. The simulation results show that compared with the traditional separated PID control method for sails and rudder, the proposed method can significantly enhance the waypoint tracking performance of unmanned sailboats under dynamic changes in wind speed and direction, and shorten the overall completion time of the waypoint tracking task. Conclusion This method can provide reliable theoretical support for unmanned sailing ships in the field of waypoint tracking control.