Leveraging Embodied Intelligence for Estimating Grasping Forces and Poses in Soft Robotic Fingers Using Deep Learning

Abstract Soft fingers, exemplifying embodied intelligence, offer significant advantages in robotic grasping due to their inherent adaptability to diverse object geometries and surface textures. This adaptability is driven by their physical structure, enabling intelligent grasping behavior without relying on complex control algorithms. By integrating artificial intelligence algorithms, this study enhances the inherent embodied intelligence of soft fingers to achieve embodied artificial intelligence (EAI), addressing the critical challenge of accurately predicting grasping forces and poses. We first conducted external interaction modeling and stability metric analysis for soft fingers. Then, we introduced morphology-driven models to predict grasping forces to achieve internal interaction modeling. Furthermore, based on the unique characteristics of soft fingers, we utilize PointNet to detect and adjust grasp poses from three-dimensional points in unstructured environments, achieving stable grasping for soft robotic systems. Our experiments, conducted on a FANUC robotic platform equipped with soft fingers, utilized force sensors to validate the grasping force prediction model and evaluated grasp pose prediction accuracy in a point cloud-based single-object environment, demonstrating a high success rate for both force prediction and grasping success. These advancements, achieved through the integration of artificial intelligence algorithms with soft fingers, validate the implementation of EAI in robotic systems, enhancing adaptability, precision, and efficiency in controlled environments, and laying the groundwork for future applications in dynamic and unstructured settings.