Intelligent design assistance for personalized medical surgery based on concentric tube continuum robots

连续体机器人在外科手术中有很大的应用潜力，因为它们细长的操纵器允许在身体深处灵活地操作，但稳定和准确。然而，它们固有的无限自由度给连续体机器人的设计带来了挑战。此外，为了找到适合个体患者的最佳手术方案，必须同时考虑机械臂参数设计和运动规划与控制。本课题开发了一种用于脑深部神经外科的同心管连续机器人的设计辅助工具。遵循系统工程中的系统设计方法，我们的方法基本上是基于1)机器人行为的精确模型和2)数学优化。在模型生成中，采用物理知识学习自动包含实验数据和专家知识，以克服经典物理模型在精度或计算负荷方面的限制。考虑到模块化结构，物理子模型以及不同的基于数据的模型，例如神经网络或通过稀疏回归的符号系统表示都可以组合在一起。这为新应用程序或新任务的快速适应提供了灵活性。机电一体化设计与运动控制的结合问题具有多准则的性质，因此将采用多目标优化的元启发式方法来解决。基于物理/数据的混合子模型约束了优化问题。此外，允许控制集是由机器学习中的分类技术生成的，以表示稳定性分析的结果。连续机器人的设计问题取决于个体患者的手术需求。因此，解决代表性病理的多目标优化问题的综合数据生成了一个数据库，可以在该数据库上学习代理模型。这个人工智能设计助手近似于多目标优化的重复解决方案，允许与工程师和医生进行实时交互。

Continuum robots have great potential in surgical applications since their slender-like manipulators allow to operate flexible --yet stable and accurate-- deep within the body.However, their intrinsically given, infinite degree of freedom has made the continuum robot design a challenge.Moreover, the design of mechanical manipulator parameters and motion planning and control have to be considered simultaneously in order to find the optimal surgical procedure for an individual patient.This project develops a design assistant tool for concentric tube continuum robots to be used in deep-brain neurosurgery.Following the systematic design methodology in systems engineering, our approach is fundamentally based on 1) accurate models of the robots behavior and 2) mathematical optimization.For model generation, physics-informed learning is used to automatically include experimental data as well as expert's knowledge to overcome limits of classical physical models in accuracy or computational load.Considering a modularized structure, both physical submodels as well as different data-based models, e.g.\ neural networks or symbolic system representations by sparse regression can be combined. This offers flexibility regarding fast adaption in case of new applications or tasks.The combined mechatronic design and motion control problem is of multicriterial nature and, thus, will be addressed by meta-heuristics of multi-objective optimization.The hybrid physics-/data-based submodels constrain the optimization problem.Additionally, the admissible control set is generated by classification techniques from machine learning in order to represent results from stability analysis.The continuum robot design problem depends on the individual patient's surgery needs.Thus, synthetic data from solving the multi-objective optimization problem for representative pathologies generates a data base on which a surrogate model can be learned.Approximating the repetitive solutions of multi-objective optimization, this AI design assistant allows for real-time interaction with engineers and physicians.