Soft Robotics is a relatively new area of research, where progress in material science has powered the next generation of robots, exhibiting biological-like properties such as soft/elastic tissues, compliance, resilience and more besides. One of the issues when employing soft robotics technologies is the soft nature of the interactions arising between the robot and its environment. These interactions are complex, and the their dynamics are non-linear and hard to capture with known models. In this thesis we argue that complex soft interactions can actually be beneficial to the robot, and give rise to rich stimuli which can be used for the resolution of robot tasks. We further argue that the usefulness of these interactions depends on statistical regularities, or structure, that appear in the stimuli. To this end, robots should appropriately employ their morphology and their actions, to influence the system-environment interactions such that structure can arise in the stimuli. In this thesis we show that learning processes can be used to perform such a task. Following this rationale, this thesis proposes and supports the theory of Soft Morphological Computation (SoMComp), by which a soft robot should appropriately condition, or ‘affect’, the soft interactions to improve the quality of the physical stimuli arising from it. SoMComp is composed of four main principles, i.e.: Soft Proprioception, Soft Sensing, Soft Morphology and Soft Actuation. Each of these principles is explored in the context of haptic object recognition or object handling in soft robots. Finally, this thesis provides an overview of this research and its future directions.