Synergistic Performance Enhancement in Heterogeneous Robot Teams

**这项研究解决了一个可能对自主机器人团队的应用产生革命性影响的基本问题：在现实世界的条件下，一个由不同种类的机器人组成的团队如何通过向同行、人员和环境学习来协同提高他们的表现？尽管多机器人团队的研究和开发取得了进展，但一个关键的挑战仍然是如何开发有效的机制，使机器人能够在现实生活中自主地产生、适应和增强群体行为，同时提高它们的个人表现。这项研究的目标是通过扩大社会行为在现实世界条件下的不同机器人团队中可以发挥的作用，以及通过研究和开发提高群体绩效的关键方面来促进自主性。它试图通过适当的机制理解和实施人的能力，学习如何在不确定和无组织的情况下，通过不仅提高自己的表现，而且通过有效的建议和批评，提高自己向他人传授知识和经验的能力，从而增加对团队合作的贡献。研究的前提是，这种计算机制在现实生活中是有效的，主要是通过将所有学习层协同集成在模块化、可扩展和分层的控制体系结构中，允许动态分布处理、通信、任务分配和协调以及故障恢复。通过利用这种体系结构，提出了一种多层方法来解决团队学习分解问题。该方法将考虑集体(个体)、协作(任务分配和协调)和协作(建议共享)三层学习行为，并尝试开发分布式马尔可夫决策过程作为这三层的学习问题模型。此外，人类顾问以来自个人和/或匿名人群的标记和未标记的训练数据的形式对提高群体绩效的贡献将通过增强的人类建议层来考虑。最后，研究了状态不确定性和团队/任务重构问题，并利用基于粒子滤波的状态估计器对已开发的并发马尔可夫决策过程进行了增强，形成了一种健壮的群体性能增强机制。其目标是使机器人能够在执行各种团队任务时处理各种情况，如感知噪声、变化的通信带宽和间歇性断开以及并发的软件和硬件故障，包括基本导航和地图绘制、碰撞避免、路径规划、以及在未知杂乱环境中的对象发现、觅食和聚集。

**This research addresses a fundamental question that can have a transformative impact on the applications of autonomous robot teams: "how can a team of heterogeneous robots synergistically enhance their performance through learning from peers, people, and environment, under real-world conditions?" Despite the advancement of research and development on multi-robot teams, a key challenge still remains as to how to develop effective mechanisms that enable the robots to autonomously generate, adapt and enhance group behaviours in real-life situations, while improving their individual performance concurrently. The goal of the research is to advance autonomy by expanding upon the role that social behaviour can play in a heterogeneous team of robots, under real-world conditions, and through studying and developing the key aspects of group performance enhancement. It attempts to comprehend and implement, through proper mechanisms, the human's capability of learning how to increase their contributions towards teamwork, in uncertain and unstructured scenarios, by enhancing not only their own performance but also their ability to transfer knowledge and experience to others through effective advice and criticism. The premise of the research is that such computational mechanisms can be effective in real-life scenarios mainly through a synergistic integration of all learning layers within a modular, scalable and hierarchical control architecture, which allows for dynamically distributed processing, communication, task allocation and coordination, and failure recovery. Through the utilization of such architecture, a multi-layered approach is proposed to the team learning decomposition problem. The approach will consider three layers of collective (individual), cooperative (task allocation and coordination) and collaborative (advice sharing) learning behaviours, and attempt to develop distributed Markov decision processes as learning problem models for the three layers concurrently. Further, the contribution of human advisors to group performance enhancement, in the form of labelled and unlabeled training data from individuals and/or anonymous crowd, will be considered through an augmenting human advice layer. Finally, the subject of state uncertainty as well as team/task reconfiguration will be explored, and the developed concurrent Markov decision processes will be enhanced with a state estimator, based on a particle filter, to form a robust group performance enhancement mechanism. The objective is to enable the robots to handle situations, such as sensory noise, varying communication bandwidth and intermittent disconnections, and concurrent software and hardware failure, in performing various team tasks, including basic navigation and mapping, collision avoidance, path planning, and object discovery, foraging and herding in an unknown cluttered environment.