To address the difficulty of modeling and control of existing soft manipulators due to their small material stiffness and unstable modulus, this study proposes a method based on a recurrent neural network (RNN) for the motion modeling of a two-degree-of-freedom (2-DOF) soft manipulator with control. A motion-capture instrument was used to collect the position coordinates under different pressures and loads, and the coordinates were imported into a gated recurrent unit (GRU) neural network model for training. The accuracy of the test set reached 98.87% when the hyperparameters were adjusted to the optimal network structure. Accordingly, a mapping function for the pressure and load at the end position was constructed. Experimental results showed that the proposed method could improve the control accuracy of the manipulator by approximately 6∼8 mm and significantly reduced the difficulty of control and modeling of a soft robot.