Soft robots are appealing in a wide variety of tasks thanks to their inherent advantages in safety, compliance, and adaptability. However, accurate modelling and control of soft robots are still significantly challenging. This paper proposes a control scheme implementing an online learning strategy. The architecture is composed of (i) a data-driven model generating a feedforward signal, and (ii) a feedback controller. The latter has two roles. Firstly, it corrects the action of the feedforward controller when the tracking error increases. Secondly, it generated a learning signal to train the data-driven model, allowing for online adaptation of the feedforward signal with respect to changes in the dynamic of the system. Experimental results show that the proposed method provides better performance compared with a PID controller when applied to a trajectory following task. Furthermore, our controller is shown to be capable of online adaptation to sudden changes in the dynamics of the soft robot due to a variable payload.