Intelligent Formation Control of Nonlinear Multi-Agent Systems

该方案的主要目的是在非线性多智能体系统及其群体的智能和鲁棒编队控制方面建立一个持久的研究计划。多智能体系统编队控制的最新研究成果大多假定智能体模型简单(点质量、单积分器和双积分器)、智能体之间理想的通信链路和理想的、无约束的车载传感器。虽然这些结果在理论上是重要的，但它们在实际多智能体应用中的应用是有限的。提出的研究计划旨在弥合这一差距，并研究刚性图形成中多智能体的智能控制。此外，多智能体将包括具有各种非线性的非线性模型，以及非理想的传感器和通信模型。该研究旨在开发一种基于递归神经网络(RNN)等智能控制工具的3D空间移动代理刚性图形成控制的新框架，该工具在经典的非线性控制中具有很大的应用前景。开发的控制器将具有实时学习能力，将能够适应系统的非线性和传感器的限制，并将具有可调整的结构，以适应现场快速变化的智能体队形。由于多智能体本质上是分布式系统，所提出的控制算法在本质上是分布式的，旨在实现群体的全局目标。本研究的创新之处在于：(1)一种非线性智能体的神经网络编队控制，其运动受图的刚性约束，控制算法和神经网络结构都是新颖的。该网络结构可根据领域中相邻智能体的数量进行实时自适应，整个网络类似于一个由多个智能体组成的多层深度学习神经网络。智能控制器的自适应将基于局部性能指标，这些指标考虑了图的刚性、能量约束和主体的非线性动态。(2)容错多智能体编队，该编队对单个智能体的故障具有健壮性，并且智能体利用局部化的群拓扑来监视和检测单个智能体的故障。(3)一种新的异质多机器人试验台，将能够在不同级别进行多智能体编队控制实验，包括静态和移动传感器、地面智能体和微型飞行器。该研究计划将结合工程学和计算机科学的两个新兴领域，即多代理系统和深度学习NNS。因此，未来的多智能体编队将提高对单个智能体故障的稳健性，将具有内置的智能以适应(单个智能体和群体)不断变化的动态，并将能够处理各种非线性和通信链路缺陷。

The main goal of this proposal is to establish a long-lasting research program in intelligent and robust formation control of nonlinear multi-agent systems and their swarms. Recent results in formation control of multi-agent systems mostly assume simple agents' models (point mass, single and double integrators), ideal communication links between the agents, and ideal, unconstrained on-board sensors. While those results are important from a theoretical standpoint, their use is limited in practical multi-agent applications.The proposed research program aims to bridge that gap and to study intelligent control of multi-agents in rigid graph formations. Moreover, multi-agents will include nonlinear models with various nonlinearities, and non-ideal sensor and communication models. The proposed research aims to develop a new framework for rigid graph formation control of mobile agents in 3D space that is based on intelligent control tools such as Recurrent Neural Networks (RNNs), which showed great promise in classical nonlinear control. The developed controllers will have real-time learning capabilities, will be able to accommodate for the systems' nonlinearities and sensors' constraints, and will have adjustable structure to account for rapidly changing formations of agents in the field. Since multi-agents are inherently distributed systems, the proposed control algorithms will be distributed in nature, while aiming to achieve a global objective of the swarm.The novelty in this research program is threefold: (1) A Neural Network (NN) formation control of nonlinear agents whose movements are constrained with graph rigidity and where both the control algorithms and the NN structure are novel. The NN structure is adaptable in real-time based on number of adjacent agents in the field and the overall network resembles a deep learning NN with many layers from different agents. Adaptation of the intelligent controller will be based on local indices of performance that consider graph rigidity, energy constraints, and agent's nonlinear dynamics. (2) A fault-tolerant multi-agent formations that are robust on individual agent failures and where agents exploit the localized swarm topology in order to monitor and detect faults of individual agents. (3) A new heterogeneous multi-robot test bed that will enable experimentation with multi-agent formation control at various levels including static and mobile sensors, ground-based agents, and micro-aerial vehicles. The research program will combine two emerging areas of engineering and computer science, i.e. multi-agent systems and deep learning NNs. As a result, future multi-agent formations will improve robustness to failures of individual agents, will have built-in intelligence to adapt to changing dynamics (of individual agents and the swarm), and will be able to handle various nonlinearities and communication links imperfections.