Design Optimization for Bellow Soft Pneumatic Actuators in Shape-Matching

Abstract

Design optimization of soft actuators is essential for task-oriented applications. Models derived from analytical solutions, the Finite Element Method (FEM), or empirical characterized datasets are widely used to estimate the response of the actuators during actuation, acting as the backbone for design optimization. Faced with the trade-off between speed and accuracy, substantial challenges occur when moving from simulation to optimization due to the compliant, high degree of freedom, and high-dimensional design space of the soft-bodied robot. FEM becomes increasingly computationally expensive with increased design complexity during optimization iterations, while the data-driven modeling approach (e.g., Artificial Neural Network) consumes significant resources prior to optimization. To address the challenge of highly nonlinear and non-convex design optimization in soft robots using the black box modeling, this paper compares of Bayesian optimization (BO) algorithm and genetic algorithm (GA) with FEM and Artificial Neural Network (ANN) models. The shape-matching of a multi-legged robot (a starfish) is demonstrated as an example of a task-oriented design scenario that presents design optimization challenges of the design space scalability. The experimental results show that the bi-level BO outperforms BO with FEM by achieving 2.8 to 9.8 times smaller objective values within a certain time for low-dimensional design problems; GA with the ANN model can achieve lower objective values 3 to 18 times faster in high-dimensional design problems than bi-level BO with FEM in low-dimensional design problems.