DISTRIBUTED SENSING, CONTROL AND DECISION MAKING IN MULTIAGENT AUTONOMOUS SYSTEMS

自主智能系统将在我们未来的社会中找到重要的应用。初步申请将在以下范畴：减灾（地震、核灾难、军事战斗、海上溢油、运输基础设施崩溃、分析和协助恐怖袭击）、偏远地点的空间探索（特洛伊小行星、火星和行星轨道观测）领域的监视、情报收集和业务控制；深海勘探和机器人技术（用于海上石油勘探灾害），随后是农业、搜索和救援、制造业和自主家用机器人等大规模应用。这些自主系统将需要机器人团队快速、适当、同时向合作伙伴提供信息的行动。它们也可以是具有传感和控制能力的基于计算网络的智能代理。这将是一项社会要求，这些（半）自主操作系统将其行为背后的原因和未来计划以简明的笔记告知人类主管，以确保其安全性和社会可接受性。基于网络的软件代理已经在我们的社会中使用了一段时间。我们的社会正在经历一场信息交换革命，朝着网络化智能设备的方向发展。这些基础设施系统中的许多都是基于定义良好的离散输入和输出，这些输入和输出要么来自人工操作，要么来自低维传感器测量。然而，在高度复杂、不断变化的环境需要被快速感知、推理和采取行动的自主机器人智能方面进展甚微。在DARPA和Robocup项目中已经报告了部分结果，这些项目没有提供全面的系统方法，或者没有完全公开。只有在机器人基础设施完备的环境中才取得进展。我们还没有一组自动驾驶汽车或代理系统的方法，可以在复杂的无基础设施环境中可靠地（半）自主地运行，从而在最少的人为监督下有效地解决问题。原因是目前的智能代理技术并没有提供解决方案。具有简单计算节点的传感器网络，是为低功耗和计算资源而开发的，不能提供解决方案。它们无法在单个代理上实现高度复杂的概念抽象。由于典型的实时性和通信瓶颈，这类智能体的计算无法被低复杂度智能体的数据融合所取代。在此项目之前，多智能体分散决策理论的方法已经发展并非常成功地使用，但尚未适当地用于多复杂智能体。该项目旨在为自主合作的多智能体系统开发一种新的方法，以提高我们的合作伙伴公司和机器人行业的技术能力。该项目将在单个代理上提供关于世界建模、态势感知、学习和信息管理的抽象功能。这些功能将在多代理合作中实现高效的实时决策，并在结构不良或基础设施自由的环境中实现分散决策。这些方法将把数字计算能力与人类的概念结构联系起来，使机器人能够像人类一样，用多层的高级和低级概念来模拟世界。

Autonomous intelligent systems will find important applications in our future society. Initial applications will be in the following areas: surveillance, intelligence gathering and operational control in the areas of disaster mitigation (earthquake, nuclear catastrophe, military combat, oil-spills at sea, transport infrastructure breakdown, analysis and assistance with terrorist attacks), space exploration at remote locations (at Trojan asteroids, on Mars and in orbit observations around planets, deep underwater explorations and robotics for offshore oil exploration disasters) followed by large scale applications such as agricultural, search and rescue, manufacturing, and autonomous household robots. These autonomous system will require quick, appropriate, and at the same time informative-to-partners, actions by teams of robots. They can also be computing network based intelligent agents with sensing and control capabilities. It will be a societal requirement that these (semi-)autonomously operating systems to inform their human supervisors about the reasoning behind their actions and their future plans in concise notes for their safety and acceptability by society.Network based software agents have been in use by our society for some time. Our society is going through information exchange revolution that is developing towards networked intelligent devices. Many of these infrastructure systems are based on well defined discrete inputs and outputs either from human operators or from low dimensional sensor measurements. Little progress has however been made in robot intelligence of autonomy where high complexity, changing environment is to be sensed, reasoned about and acted upon quickly. Partial results have been reported in DARPA, Robocup projects that do not provide comprehensive systematic approach or are not fully publicly available. Progress has only been made in heavily infrastructured environments of robots. We do not yet have the methodology for a set of autonomous vehicles or agent systems to operate reliably and (semi-)autonomously in complex infrastructure-free environments to solve problems efficiently with minimal human supervision. The reason is that current intelligent agent technology does not provide solutions. Sensor networks with simple computational nodes, that were developed for low power and computational resources do not provide solutions. They miss the ability of high complexity conceptual abstractions onboard a single agent. The computations of these type of agents cannot be substituted by data fusion of low complexity agents due to typical real-time and communication bottlenecks. Methods of multi-agent decentralized decision theory have been developed and very successfully used prior to this project but have not been properly exploited for multiple complex agents.This project intends to develop a new methodology for autonomous cooperating multi-agent systems that is to boost the technological capabilities of our partner companies and the robotics industry in general. The project will provide the missing capabilities of abstractions concerning world modeling, situational awareness, learning and information management onboard a single agent. These capabilities will enable efficient realtime decision making within multi-agent cooperation and decentralized decision making in poorly structured or infrastructure free environments. These methods will connect digital computing power with human conceptual structures to enable robots to model the world with layers of high and low level concepts as humans do.