Supporting files for thesis "Deep-learning-based Morphological Modelling: Case Study in Soft Robot Control, Shape Sensing and Deformation"

The main focus of this thesis is to develop appropriate morphology modelling strategies for typical deformable structures by integrating physics-aligned prior knowledge and deep learning algorithms. To perform accurate control for soft continuum robots, a reinforcement-learning (RL)-based framework is proposed, which integrates conventional piecewise curvature constant (PCC) model and adaptive learning strategies. The algorithm of deep deterministic policy gradient (DDPG) along with domain randomization and offline retraining facilitates fast initialization and stable path following, even under varying tip load, demonstrating its advantages over Jacobian model-based and supervised-learning-based control methods. Not only tube-like soft robots, this thesis also endeavors to enhance the proprioception capabilities of soft flexible surfaces. A real-time shape sensing framework is proposed that combines finite element analysis (FEA) with autoregressive deep learning, requiring only sparsely distributed sensor nodes to predict continuous high-order complex 3D shapes. The cost-effective optical waveguide is utilized to detect the light loss induced by surface deformation. While sharing similarities with thin surfaces in terms of structure, cloth presents unique challenges due to its hyper flexibility and lack of adhesive sensing capabilities. To achieve real-time shape prediction of the cloth when in contact with human bodies or obstacles, a framework based on graph-neural-networks (GNNs) is introduced. Leveraging graph structure and spiral graph convolution, this approach facilitates efficient processing of high-dimensional cloth data while preserving essential geometry information. Additionally, a real fabric motion capture system is developed to establish a real 3D cloth library, aiding research on fabric deformation. Considering 2D images are more accessible and common compared to 3D topology data, this thesis also investigates soft tissue modelling via medical images. To explore the hidden tissue properties through ultrasound B-mode images, a framework that incorporates U-Net and ultrasound propagation physics is proposed. The framework can estimate physics-aligned tissue properties which characterize ultrasound attenuation, reflection and backscattering, thereby showing potential for enhancing soft tissue deformation modelling.本论文的核心研究方向为：融合物理对齐先验知识与深度学习算法，为典型可变形结构开发适配的形态学建模方案。为实现软体连续体机器人的精准控制，本文提出一种基于强化学习（Reinforcement Learning, RL）的框架，该框架整合了传统分段曲率恒定（Piecewise Curvature Constant, PCC）模型与自适应学习策略。深度确定性策略梯度（Deep Deterministic Policy Gradient, DDPG）算法结合域随机化与离线重训练技术，可实现快速初始化与稳定的路径跟踪，即便在末端负载变化的场景下仍表现优异，其性能优于基于雅可比模型与监督学习的控制方法。除管状软体机器人外，本论文还致力于提升柔性表面的本体感知能力。本文提出一种实时形状感知框架，将有限元分析（Finite Element Analysis, FEA）与自回归深度学习相结合，仅需稀疏分布的传感器节点即可预测连续高阶复杂三维形状。研究采用高性价比光波导检测由表面变形引发的光损耗。尽管布料在结构上与薄型表面存在相似性，但因其超高柔韧性且缺乏粘附传感能力，仍面临独特的挑战。为实现布料与人体或障碍物接触时的实时形状预测，本文提出一种基于图神经网络（Graph Neural Networks, GNNs）的框架。该方法借助图结构与螺旋图卷积，可高效处理高维布料数据的同时保留关键几何信息。此外，本文搭建了真实布料运动捕捉系统以构建真实三维布料数据集，为布料变形相关研究提供支撑。相较于三维拓扑数据，二维图像的获取更为便捷且普及，因此本论文还探索了基于医学图像的软组织建模方法。为通过超声B型图像挖掘组织的潜在特性，本文提出一种融合U-Net与超声传播物理的框架。该框架可估算符合物理对齐的组织特性，这些特性表征了超声衰减、反射与背散射行为，进而在提升软组织变形建模能力方面展现出应用潜力。