4D printing soft robots guided by machine learning and finite element models

Robotics has been synonymous with high precision and rigidity. However, in recent years the new technology of soft robotics has emerged adding flexibility and adaptability that was not previously possible with rigid robots. The advancements of 3D printing manifested in soft robotics as 4D printing, where the fourth dimension refers to the time-dependent response of the printed mechanism to variable stimuli, such as heat, electricity, magnetism, and pneumatic pressure.

The 4D printed soft pneumatic actuator (SPA) focused on this study consists of finger-like structures with bellows, which inflate when pressurized causing it to extend and bend. Compared to the rigid counterparts, it is low cost, lightweight, easy to manufacture, adaptable, flexible, and deformable with applications in medical devices for surgery, therapy, and rehabilitation, and industrial grippers for fruit and food picking and placing.

3D/4D printing is being studied due to its capability to produce soft robots and actuators with complex inner structures. Ninjaflex material proved to be a good choice as it could be 3D printed without air bubbles and has hyperelastic properties that provide flexibility and sensitivity to applied stress. One of the current challenges in the 4D printing of soft robots and actuators is the modeling and prediction of their motion particularly due to the nonlinearity of the material. A linear analytical model often fails to accurately predict the actuation behavior of the 4D printed actuators whereas numerical simulations incorporated with nonlinear material principles improve accuracy.

However, using the advantage of machine learning (ML) techniques trained via numerical results could help save time and effort during the design. A pure data-driven modeling approach enables the prediction of actuation behavior under different operating conditions without precise material models. A combination of the accuracy of FEM modeling and the time efficiency of ML to formulate a neural network algorithm was introduced. The FEM was employed for training the data and the artificial neural network (ANN) to create the model, a method that will be utilized in this study to predict the deflection angle in response to a varying input air pressure.

Having achieved a reliable ML model to predict the bending angle of actuators the study continued to estimate the geometrical specifications required to develop 4D-printed SPAs based on the input variables. The classification model for predicting shape is useful for design purposes and enables the user to achieve their desired motion in 4D printing if, for example, there are geometrical requirements or limitations. The results from the one-way analysis of variance (ANOVA) enabled the identification of the variables that will maximize actuation which can ultimately lead to efficiency in energy requirements and cost of the actuator.

The classification model was developed by optimizing the hyperparameters to create a model with 94.3 % accuracy of distinguishing between the three actuator shapes and the results were experimentally validated on grasping different size objects. This approach of ML-based modeling could be tailored to the other types of 4D printing and 3D/4D-printed soft robotics studies where the new design parameters could be defined first and after validating the experimental results in FEM all the simulation data could be used to create a new ML model. This work may be used to develop control systems utilizing this technology and design, or for design purposes of closed-loop 4D-printed soft robots and actuators with specific functional and geometrical requirements.