Kinematics-Informed Neural Networks: Enhancing Generalization Performance of Soft Robot Model Identification

Abstract

A hybrid system combining rigid and soft robots (e.g., soft fingers attached to a rigid arm) ensures safe and dexterous interaction with humans. Nevertheless, modeling complex movements involving both soft and rigid robots presents a challenge. Additionally, the difficulty of obtaining large datasets for soft robots, due to the risk of damage by repetitive and extreme actuations, hiders the utilization of data-driven approaches. In this study, we present a Kinematics-Informed Neural Network (KINN), which incorporates rigid body kinematics as an inductive bias to enhance sample efficiency and provide holistic control for the hybrid system. The model identification performance of the proposed method is extensively evaluated in simulated and real-world environments using pneumatic and tendon-driven soft robots. The evaluation result shows employing a kinematic prior leads to an 80.84% decrease in positional error measured in the L1-norm for extrapolation tasks in real-world tendon-driven soft robots. We also demonstrate the dexterous and holistic control of the rigid arm with soft fingers by opening bottles and painting letters.