CAREER: MaLPhySiCS - Machine Learning-assisted Physics-based Simulation and Control of Soft robots

该学院早期职业发展（CAREER）资助将支持制定和实验验证用于软机器人建模的数值工具的研究。软机器人通常是通过涉及几个原型的艰苦的试错过程来设计和控制的。该项目旨在使用实时物理引擎自动化此设计和控制过程。该仿真解决了机器人仿真中的几个关键挑战：（1）大的结构变形，（2）与水动力的非平凡耦合，以及（3）接触和碰撞的优势。模型降阶是将复杂机器人模型转化为仿真框架的必要条件。机器学习（ML）提供了一套特殊的工具，可以将复杂的结构简化为模型，而不会影响准确性。基于物理和ML的快速仿真工具将用于构建细菌启发的软机器人的控制框架。该项目的研究目标是制定快速有效的基于物理的仿真，并辅以机器学习，用于软机器人的自主控制。使用这个框架，一个大规模的细菌启发的机器人将被设计和控制。该机器人将利用鞭毛（薄而灵活的尾巴）的屈曲来控制其游泳方向。这有望成为具有单一标量控制输入的最简单的自主软机器人。在该项目中要解决的两个关键挑战是：（1）计算效率，使模拟可以用于优化，和（2）物理精度和鲁棒性的模型，使基于模型的控制可以采用在真实的机器人系统。为了实现这一目标，计划在基于离散微分几何的仿真框架中对复杂系统进行机器学习辅助建模。将开发机器人结构和流体动力学的神经网络模型。这些模型有望与简化的启发式模型一样快，与基于物理的细粒度模型一样准确。该仿真工具将被用来开发一个基于模型的控制框架的细菌启发的机器人不受约束的自主操作。这个机器人可以帮助我们深入了解细菌的运动，例如，不稳定性在细菌推进中的作用从机器人的角度来看，机器人只有一个控制输入，移动部件的数量很少。该机器人的设计使其适合小型化到亚毫米级，具有潜在的生物医学应用。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

This Faculty Early Career Development (CAREER) grant will support research that will formulate and experimentally validate numerical tools for modeling of soft robots. Soft robots are typically designed and controlled through a painstaking trial-and-error process involving several prototypes. This project seeks to automate this design and control process using a real-time physics-engine. The simulation addresses several key challenges in simulation of robots: (1) large structural deformation, (2) nontrivial coupling with hydrodynamic forces, and (3) preponderance of contact and collision. Model reduction is necessary to cast a complex robot into the simulation framework. Machine learning (ML) provides an exceptional set of tools to reduce complex structures into a model without compromising accuracy. The fast simulation tool, based on physics and ML, will be used to build a control framework for a bacteria-inspired soft robot.The research objective of this project is formulation of fast and efficient physics-based simulation, assisted by machine learning, for autonomous control of soft robots. Using this framework, a macroscale bacteria-inspired robot will be designed and controlled. This robot will use buckling in flagellum (thin flexible tail) to control its swimming direction. This is expected to be the simplest autonomous soft robot with a single scalar control input. Two key challenges to be tackled in the project are: (1) computational efficiency so that the simulation can be used for optimization, and (2) physical accuracy and robustness of the models so that model-based control can be employed on the real robotic systems. Towards this goal, machine learning-assisted modeling of complex systems in a discrete differential geometry-based simulation framework is planned. Neural network-based models for the structure of the robot and the hydrodynamics will be developed. These models are expected to be as fast as simplified heuristic models and as accurate as physics-based fine-grained models. This simulation tool will be used to develop a model-based control framework for the bacteria-inspired robot for untethered autonomous operation. This robot can help us gain insight into bacterial locomotion, e.g., role of instability in bacterial propulsion. From a robotics perspective, the robot has only one control input with minuscule number of moving parts. The design of the robot makes it amenable for miniaturization to sub-millimeter scale with potential biomedical applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.