Bayesian Optimization of Pneumatic Soft Grippers via Reconfigurable Modular Molds

Design optimization of soft grippers is critical to functionally exploit their compliance. However, it is difficult to predictably model or search the design space of even simple Fluidic Elastomer Actuators. This work presents a method to rapidly customize and identify desirable morphologies of pneumatic soft grippers via Bayesian Optimization of reconfigurable modular molds. With the goal of maximizing grasping success rate for a general object set, the reality-assisted optimization process uses results from physical pick and place experiments to iterate through a large array of design parameters. Suggested design parameters dictate the assembly of pre-fabricated modules in the mold to generate silicone-casted soft fingers. These fingers are integrated to form a gripper and tested to inform the next iteration. This process allows faster iterations compared to 3D printing equivalent grippers or molds, and discretizes the design space for faster search through parameter combinations. An improvement of 34% in average grasping success rate was achieved in six iterations, shedding light on desirable parameter configurations for the grasping task.