Rapid fault-recovery strategies for resilient robot swarms

机器人正越来越多地成为我们日常生活中重要的一部分，它自动完成一些任务，比如保持我们的家清洁，以及在大型仓库挑选/打包我们的包裹。人口老龄化和在危险和重复的任务中取代人类工人的需求现在导致了新的任务即将出现(例如，在农业自动化和环境监测方面)，要求我们的机器人做更多的事情，作为群体(一大群机器人)的一部分进行大量工作，在广阔的区域内协调感知和行动，并有效地执行它们的任务。然而，到目前为止，我们的机器人群还没有为部署做好准备；它们无法处理在运行过程中不可避免的损坏和故障，它们仍然是脆弱的系统，在困难的条件下无法发挥作用。这个项目的目标是通过开发算法来纠正这种情况，使机器人群在不超过几分钟的时间内迅速从群中机器人遭受的故障和损坏中恢复。现有的机器人群容错系统是有限的。它们被限制为只能诊断设计者先验预测的故障，这几乎不能涵盖机器人群在复杂环境中长时间运行时可能遇到的所有可能情况。群体中的机器人数量众多，它们可以相互作用的方式也非常复杂，这使得预测潜在故障和预先定义相应的恢复策略变得困难；这可能解释了为什么现有的故障检测和故障诊断系统都没有扩展到为机器人群体提供故障恢复机制。因此，为了设计机器人群的容错算法，我们需要超越传统的依赖故障诊断信息进行故障恢复的方法，而是将机器人群中的故障恢复描述为一个在线行为适应过程。有了这样的方法，群体中的机器人通过反复试验学习尽管有故障仍能工作的新的补偿行为，从而使它们的行为适应持续的故障。然而，目前学习新的机器人群体行为的方法是耗时的，需要几个小时。因此，这种方法不适合用于快速故障恢复的行为适应(学习新的群体行为)，为了有效的故障恢复，行为适应要求机器人群创造性地、快速地在线学习新的补偿性群体行为，在持续故障的情况下工作，有效地从故障中恢复群体。该提案将通过研究快速在线行为适应的数据高效机器学习技术来满足这些要求，该技术以创造性的和自动产生的直觉--离线进化--为指导，对工作群体行为进行指导。由此产生的系统将对机器人群的长期运营产生重大影响，并为它们的部署开辟新的有趣的应用程序，例如使用一群自动水面车辆监测大片水体的污染物。

Robots are increasingly becoming an important part of our day-to-day lives, automating tasks such as keeping our homes clean, and picking/packing our parcels at large warehouses. An aging population and the need to substitute human workers in dangerous and repetitive tasks have now resulted in new tasks on the horizon (e.g., in agriculture automation and environmental monitoring), requiring our robots to do more, to work in large-numbers as part of a swarm (a large team of robots), to coordinately sense and act over vast areas, and efficiently perform their mission. However, our robot swarms to date are unprepared for deployment; unable to deal with the inevitable damages and faults sustained during operation, they remain frail systems that cease functioning in difficult conditions. The goal of this project is to remedy this situation by developing algorithms for robot swarms to rapidly -- in no more than a few minutes -- recover from faults and damages sustained by robots of the swarm.The existing fault-tolerant systems for robot swarms are limited. They are constrained to only diagnose faults anticipated a priori by the designer, which can hardly encompass all the possible scenarios a robot swarm may encounter while operating in complex environments for extended periods of time. The multitude of robots in a swarm and the large number of intricate ways they can interact with each other makes it difficult to predict potential faults and predefine corresponding recovery strategies; which may explain why none of the existing fault-detection and fault-diagnosis systems have been extended to provide fault-recovery mechanisms for robot swarms. Therefore, in order to design fault-tolerant algorithms for robot swarms, we need to move beyond the traditional approaches relying on fault-diagnosis information for fault recovery.Fault recovery in a robot swarm may instead be formulated as an online behavior-adaptation process. With such an approach, the robots of the swarm adapt their behavior to sustained faults by learning via trial-and-error new compensatory behaviors that work despite the faults. However, the current approaches to learning new robot swarm behaviors are time-consuming, requiring several hours. Therefore, such approaches are inappropriate for behavior adaptation (learning new swarm behaviors) for rapid fault recovery.Behavior adaptation for effective fault recovery requires the robot swarm to creatively and rapidly learn new compensatory swarm behaviors online, that work despite the sustained faults, effectively recovering the swarm from the faults. The proposal will address these requirements by investigating data-efficient machine learning techniques for rapid online behavior adaptation, guided by creatively and automatically generated intuitions -- evolved offline -- of working swarm behaviors. The resulting system would have a significant impact on long-term operations of robot swarms, and open up new and interesting applications for their deployment, such as the monitoring of large bodies of water for pollutants using a swarm of autonomous surface vehicles.