Abstract
Soft robots are a major type of robotic systems characterized by continuous deformation, high compliance, and the ability to adapt to unstructured environments. This morphology offers unique advantages for adaptable and safe interactions and operations in dynamic environments. Soft robots have theoretically infinite degrees of freedom (DOFs) and can exhibit highly nonlinear deformation behavior, which pose challenges in accurate modeling and real-time control. Remarkable modeling methods have been performed, ranging from physics-based analytical and numerical methods to reduced-order methods and data-driven methods using machine learning (ML) techniques, to capture the complex dynamics and kinematics of soft continuum structures. To enable generalizable modeling, the goal of this dissertation is to establish a physics-guided deep learning (PGDL) framework for soft robots’ modeling and learning, which can adapt to unseen/unknown conditions by integrating first-principles physics into deep learning.
This dissertation first investigates soft robot fabrication, comparing traditional silicone molding with fused deposition modeling (FDM), where FDM-based 3D printing proves to be a more efficient alternative, especially for complex geometries. Using this approach, a novel 3D-printed soft robotic glove with a pneumatic actuation mechanism was designed for rehabilitation and assistive robotics. The developed hand exoskeleton was successfully integrated with a control system, enabling adaptive grasping and dexterous manipulation, assisting individuals with disabilities in essential daily tasks. Using 3D-printed soft robotics as testbeds, we developed a physics-guided deep learning (PGDL) framework that integrates first-principles physics with deep learning as a surrogate model. The framework employs physics informed neural networks (PINNs) to solve partial differential equations (PDEs) in spatial coordinates, using the Navier-Cauchy equation as the governing equation. The stress field is derived from stress-strain- displacement relationships to ensure consistency with physical deformation. To further consider the intrinsic reasons for the nonlinearity of soft robotics in modeling to ensure accuracy and generability, we extended the proposed PGDL framework to account for the geometry and material nonlinearity of soft robot mechanics into PINN, namely Soft-PINN. For geometry nonlinearity, which arises from substantial configuration changes, especially during large deformation, our method addresses it through an updated spatial geometry strategy, borrowing the idea of Lagrangian-Eulerian (ALE) finite element formulations (FEFs). For material nonlinearity, we addressed it by embedding hyperelastic material models into the Soft-PINN framework via a strain energy function derived from the first and second invariants of the Green deformation tensor. Together, these enhancements enable the proposed PINN architecture to robustly simulate highly deformable soft structures across varying geometries and loading conditions. In real-world applications, such as grasping and manipulation, we introduce a soft-contact mechanics component to model the interaction between deformable soft robot structures and unknown objects, enabling material-awareness soft grasping. Our method embeds the Hertzian-based contact modeling as the boundary condition, along with the related physical hyperelastic model, as the governing equation within the PDE formulation. This dissertation research is among the first to integrate fundamental physics into deep learning as a generalizable technique for modeling soft continuum robots.
