CPS: Medium: Collaborative Research: Towards optimal robot locomotion in fluids through physics-informed learning with distributed sensing

鱼类是流体运动的大师，因为它们高度集成了生物传感，计算和运动系统。它们善于从周围的液体中收集和利用丰富的信息，用于水下传感和运动控制。受鱼类游泳的启发和通知，这项研究的目的是开发一种新的生物启发的网络物理系统（CPS），集成？身体上的机器鱼与流体环境的结合？网络？机器人控制机器学习算法。具体地，该CPS系统包括i）具有分布式感测能力以收集流量信息的压力感测皮肤，ii）计算由接收压力感测反馈的中央模式发生器（CPG）输出的机器人马达信号的控制和学习算法，iii）实现和验证用于水下感测和控制任务的CPS框架的机器鱼平台，以及iv）用于调查和模拟基本流体物理的实验和计算方法。该CPS系统将对核心CPS研究领域产生直接影响，如设计，控制，数据分析，自治和实时系统。它还将对广泛的工程应用产生重大影响，这些应用需要分布式传感，控制和自适应驱动。例子包括人机交互，医疗机器人，无人驾驶航空器/水下航行器，药物给药，医疗治疗和空间可部署结构等。利用这项研究的多学科性质，该奖项将支持各种教育和推广活动。特别是，将开展一系列扩大参与工程学的活动。 该研究项目集成了多种CPS技术，以开发用于鱼群控制的生物启发技术。其中包括压力敏感皮肤项目中的inthanovations，该项目将首先开发一种分布式压力敏感合成皮肤，该皮肤将安装在机器鱼上，以绘制其身体和尾鳍表面的压力分布。分布的压力信息，然后将用于反馈控制策略，调制CPG产生尾鳍运动模式的机器鱼。控制策略和尾鳍运动模式将通过强化学习优化，首先在替代流体环境中，然后在真正的流体环境中。替代流体环境将使用数据驱动的非参数模型进行开发，这些模型由基于物理的鱼类游泳流体动力学模型提供信息，并使用实验和计算流体动力学（CFD）模拟数据进行训练。上述控制学习方法也将用于实现一组机器鱼的有效集群，这些机器鱼由CPG单独控制，通过周围的流体和压力感觉反馈相互作用。将使用CFD模拟研究机器鱼的优化游泳/集群性能和基本物理。总之，这项研究将推进CPS知识：1）设计和创造的电子和传感器材料和设备的机器人皮肤应用; 2）为在复杂环境中运行的机器人系统开发数据高效，物理信息的学习方法，特别是利用深度学习的最新进展，利用水下传感和机器人压力数据的空间和时间丰富性对照组;以及3）水流物理学和鱼类游泳的建模。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Fishes are masters of locomotion in fluids owing to their highly integrated biological sensing, computing and motor systems. They are adept at collecting and exploiting rich information from the surrounding fluids for underwater sensing and locomotion control. Inspired and informed by fish swimming, this research aims to develop a novel bio-inspired cyber-physical system (CPS) that integrates the ?physical? robot fish and fluid environment with the ?cyber? robot control & machine learning algorithms. Specifically, this CPS system includes i) a pressure sensory skin with distributed sensing capability to collect flow information, ii) control and learning algorithms that compute robot motor signals, output by central pattern generators (CPGs) which receive pressure sensory feedback, iii) a robot fish platform to implement and validate the CPS framework for underwater sensing and control tasks, and iv) experimental and computational methods to investigate and model the underlying fluid physics. This CPS system will have immediate impacts on the core CPS research areas such as design, control, data analytics, autonomy, and real-time systems. It will also significantly impact a wide range of engineering applications which demand distributed sensing, control and adaptive actuation. Examples include human-machine interactions, medical robots, unmanned aerial/underwater vehicles, drug dosing, medical therapeutics, and space deployable structures among others. Leveraging the multidisciplinary nature of this research, this award will support a variety of educational and outreach activities. In particular, a list of activities in broadening participation in engineering will be carried out. This research project integrates multiple CPS technologies to develop bio-inspired technologies for swarm control of fish. These include inthanovations in a pressure sensitive skin project will first develop a distributed pressure sensitive synthetic skin, which will be installed on robotic fishes to map the pressure distribution on their body and caudal-fin surfaces. The distributed pressure information will then be used in a feedback control policy that modulates CPGs to produce caudal-fin motion patterns of the robotic fishes. The control policy and the caudal-fin motion patterns will be optimized via reinforcement learning first in a surrogate fluid environment and then in the true fluid environment. The surrogate fluid environment will be developed using data-driven non-parametric models informed by physics-based hydrodynamic models of fish swimming, trained using combined experimental and Computational Fluid Dynamics (CFD) simulation data. The above control-learning methods will also be used to achieve efficient schooling in a group of robotic fishes, individually controlled by a CPG, which interacts with each other through surrounding fluids and pressure sensory feedback. The optimized swimming/schooling performance of robotic fishes and the underlying physics will be studied using CFD simulation. Together, this research will advance CPS knowledge on: 1) the design and creation of electronic and sensor materials and devices for robot skin applications; 2) the development of data-efficient, physics-informed learning methods for robotic systems that operate in complex environments, especially leveraging the recent progress on deep learning to exploit the spatial and temporal richness of the pressure data for underwater sensing and robot control; and 3) the flow physics and modeling of fish swimming.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.