Application of Bayesian Optimization in Gripper Design for Effective Grasping

Despite many recent technological advancements, grasping remains a challenging open problem in robotic manipulation. In contrast with most research which focuses equipping grippers with varying degree of intelligence, we approach grasping from a gripper design perspective, aiming to find the best tool for grasping a specific set of objects. Building on our previous work, this paper reviews a suitable parametrization for the geometry of two common families of industrial grippers and presents a grasp score beneficial for gripper design. We then formally cast the problem of finding the best gripper parametrization within a probabilistic framework, addressing it using Bayesian Optimization tools. Numerical results on a set of industrial objects demonstrate the effectiveness of the approach showing up to $\approx 300 \%$ improvement compared to the performance obtained using a fixed set of grippers.