Modeling of Soft Robotic Grippers for Reinforcement Learning-based Grasp Planning in Simulation

Most grasping solutions in the literature and industry rely on learning-based planners developed for grippers with rigid fingers, whose grasp geometries can be abstracted deterministically into simple shapes, typically in terms of a single grip width parameter. Soft grippers, on the other hand, have nonlinear relationships between the actuation input and final geometric shape of the grasping surface. Modeling this relationship is important for training accurate reinforcement learning-based grasp planners. In this paper, we present a prototype cable-driven soft robotic gripper, and describe a computer vision-based technique with LASSO regression to transfer the relationship between the length of the cable and the finger's shape parameters in terms of angular deflections to a PyBullet simulation environment. The average root mean-squared error for this regression model was 0.15 rad. This work forms the first step of a proposed real-to-sim-to-real pipeline for training physically accurate soft robotic grasp planners.