Soft robots must be able to structure an automation problem into a sequence of actions that lead to a desired state, before they can fulfill a meaningful role in automation applications. This, however, can only be successful if the robot can predict the outcome of an action. The theory of rigid industrial robots is not applicable without major changes, because kinematic chains do not adequately describe the continuous deformation of the complex, often biologically inspired shapes of soft robots. Analytic solutions have not been found yet. Numerical solutions based on finite elements are slow, technically challenging, and only suitable for one specific robot. It is, however, possible to observe the outcome of an action, and use these observations to plan a sequence of actions that let the robot accomplish an automation task. In this paper, we analyze a probabilistic automaton that computes the optimal sequence of actions to bring the robot into a desired state. An earlier article explained the functioning of the method in a toy example. In this paper, we analyze if it is feasible to apply the method to a planning problem inspired by a real soft robot. We show the results and document the planning process. We identify the analog of an impulse response, although it is not closed form due to the nonparametric nature of the method. Note to Practitioners —A soft robotic sorting table has a computer-controlled soft surface that can move delicate objects without damaging them. There are currently no closed-loop control systems for such robots, because it is unclear how to relate the control signals to the behavior of the table, or which actions to choose in order to solve a manipulation task. In this paper, we propose a probabilistic automaton to plan the best action sequence on average. The sequence brings the workpieces on top of a soft table into a desired condition. It is a machine learning solution that is based on observations of the input signals and their effect, rather than a detailed analytical or numerical modeling of the robot. We show that it is feasible to model an existing soft robotic table. We demonstrate that the planning is successful by solving complex maze tasks. Our results are based on experiments and simulations.