A Meta-Model for The Design of Soft Pneumatic Actuators Using Neural Networks and Finite Element Analysis

Abstract

Previous works have demonstrated that complex soft robots can be built from simple building blocks, as evidenced by studies using no more than four primitive units to develop locomoting robots. However, these studies were restricted to idealized, non-physical deformations in physics engines lacking real-world actuation like pneumatic actuation. This study addresses this gap by utilizing non-linear finite element simulations and a custom Moore neighborhood encoding to model the deformation of primitive units for pneumatic actuation. A neural network meta-model is trained on a large dataset of automated simulations, allowing efficient soft robot design. The efficacy of the encoding and meta-model is demonstrated through the design of a pneumatically actuated asymmetric bending actuator. This design, though surprisingly different from conventional actuators, demonstrates high effectiveness. The meta-model's computational efficiency enables five optimization restarts in under 3% of the time required for a single finite element simulation, highlighting the encoding's ability to efficiently explore the design space and create high-performance soft robots.