Bayesian design optimization of biomimetic soft actuators

Abstract

The design of versatile soft actuators remains a challenging task, as it is a complex trade-off between robotic adaptability and structural complexity. Recently, researchers have used statistical and physical models to simulate the mechanical behavior of soft actuators. These simulations can help identify optimal actuator designs that fulfill specific robotic objectives. However, automated optimization of soft robots is a delicate balance between simplifying assumptions that reduce predictive fidelity and expensive simulations that limit design space exploration. Here we propose a generalized Bayesian optimization method to identify the designs of fiber-based biomimetic soft-robotic arms that minimize the actuation energy under arbitrary robotic control requirements. We use the reduced-order active filament theory as the overarching design paradigm and mechanical model, which enables a computationally robust and efficient optimization process. We evaluate the performance of our Bayesian optimization for a simple control objective in which the actuator has to reach a given target position. We show that our proposed optimization scheme outperforms a random-search baseline; it identifies more desirable designs faster and more frequently. Although we illustrate the performance of our approach for a single actuation problem, the derived method easily generalizes to the design optimization of fiber-based actuators under a large family of robotic applications.