Cross-sectional Topology Optimization of Slender Soft Pneumatic Actuators using Genetic Algorithms and Geometrically Exact Beam Models

TL;DR

The paper tackles the challenge of designing slender soft pneumatic actuators without laborious trial-and-error by applying black-box topology optimization to the actuator cross-section. A polar Voronoi discretization, a geometrically exact Cosserat beam model, and a three-chamber, three-pressure setup enable simulation-driven optimization of the cross-section to reach a defined end-effector workspace under pressures $P_i$. Through a genetic algorithm with two recombination strategies and two mutation schemes, the study demonstrates that viable cross-sectional designs can be found for three target workspaces, reducing manual design iterations, though experimental validation remains outstanding. The work lays groundwork for rapid, simulation-based design of soft actuators and points to future improvements including torsion, radial reinforcement, real-world manufacturing, and differentiable simulators to further accelerate optimization.

Abstract

The design of soft robots is still commonly driven by manual trial-and-error approaches, requiring the manufacturing of multiple physical prototypes, which in the end, is time-consuming and requires significant expertise. To reduce the number of manual interventions in this process, topology optimization can be used to assist the design process. The design is then guided by simulations and numerous prototypes can be tested in simulation rather than being evaluated through laborious experiments. To implement this simulation-driven design process, the possible design space of a slender soft pneumatic actuator is generalized to the design of the circular cross-section. We perform a black-box topology optimization using genetic algorithms to obtain a cross-sectional design of a soft pneumatic actuator that is capable of reaching a target workspace defined by the end-effector positions at different pressure values. This design method is evaluated for three different case studies and target workspaces, which were either randomly generated or specified by the operator of the design assistant. The black-box topology optimization based on genetic algorithms proves to be capable of finding good designs under given plausible target workspaces. We considered a simplified simulation model to verify the efficacy of the employed method. An experimental validation has not yet been performed. It can be concluded that the employed black-box topology optimization can assist in the design process for slender soft pneumatic actuators. It supports at searching for possible design prototypes that reach points specified by corresponding actuation pressures. This helps reduce the trial-and-error driven iterative manual design process and enables the operator to focus on prototypes that already offer a good viable solution.

Figures (7)

• Figure 1: Discretization of the cross-section with a) the slender soft pneumatic actuator from initial cylindrical shape to a sculpted shape with air chambers and b) the discretization of the cross-section of this cylindrical soft pneumatic actuator and assignment of air chambers (red, yellow, blue) and material (gray).
• Figure 2: Discretized circular cross-section of a soft pneumatic actuator generated by a random Voronoi tessellation with air chambers (red, yellow, blue) and material (gray)
• Figure 3: a) Randomized circular cross-section with the corresponding workspace of the end-effector rendering in top-down view, and b) the same cross-section and workspace rendered in a diagonal view.
• Figure 4: First case study with a random cross-section to generate target workspace. a) The randomly generated cross-section and the computed workspace as target workspace in top-down view and diagonal view, b) the best individual (cross-section and workspace) obtained through all iterations and methods and the target workspace (pink), c) the best individual (cross-section and workspace) obtained at the end of the worst performing iteration through all iterations and methods and the target workspace (pink)
• Figure 5: Second case study with a specific cross-section to generate target workspace. a) The manually designed cross-section and the computed workspace as target workspace in top-down view and diagonal view, b) the best individual (cross-section and workspace) obtained through all iterations and methods and the target workspace (pink), c) the best individual (cross-section and workspace) obtained at the end of the worst performing iteration through all iterations and methods and the target workspace (pink)