In this paper we introduce a neural network-based approach to tensegrity morphing: the task of actively changing the shape of a tensegrity structure to “fit” between obstacles in a cluttered environment. We specifically focus on a class of forming tasks, when the robot is required to pass between two parallel plate-like obstacles, and develop a robust solution both for generating dataset and for training predictor. Proposed predictor is able to predict both the shape of the tensegrity structure and the desired rest lengths of the actuated elastic elements, which can serve as motor commands when a quasi-static configuration-space trajectory tracking is used. We demonstrate high accuracy on validation dataset, and show the conditions when predictor overfits.