Soft robots made from soft materials recently attracted tremendous research owing to their unique softness compared with rigid robots, making them suitable for applications such as manipulation and locomotion. However, also due to their softness, the modeling and control of soft robots present a significant challenge because of the infinite degree of freedom. In this case, although analytic solutions can be derived for control, they are too computationally intensive for real-time application. In this paper, we aim to leverage reinforcement learning to approach the control problem. We gradually increase the complexity of the control problems to learn. We also test the effectiveness and efficiency of reinforcement learning techniques to the control of soft robots for different tasks. Simulation results show that the control commands to be computed in milliseconds, allowing effective control of soft manipulators, up to trajectory tracking.