Interactions between robots and the environment frequently occur during most modern robotic applications. In- dependently from the nature - intentional or unintentional - of these interactions, they must be properly considered during robot design and control, otherwise, they may lead to dangerous and harmful scenarios, both for robots and for surrounding people. From the design side, a typical solution is soft robotics, i.e., the addition of elastic elements into the robot structure. However, the compliant robot behavior conferred by the soft elements could be drastically altered by the the adopted control technique. In this paper, we propose a control framework based on Iterative Learning Control for articulated soft robots, which is able to cope with the complex dynamics of these systems, while achieving good tracking performance. The preservation of the robot compliance is obtained through constraints on the feedback components. The main limitations of learning approaches are the issues related to generalization and scalability. For this reason, we tackle this problem by proposing solutions to generalize the learned control action w.r.t. stiffness profile, velocity execution, and trajectory.