Coupled Navigation-Networking Techniques for Mobile Robot Systems

驾驶员辅助技术和全自动驾驶汽车目前正在实现，自动化无人机群正在寻求解决从搜索和救援到货物交付的复杂任务。这些技术在很大程度上是作为单个车辆的独立平台实施的，随着大规模部署的开始，这种策略可能会导致次优行为。使用共享数据网络的车辆之间合作的可能性是有希望的；这涉及到网络协议和多机器人控制算法之间的交互。这些相互作用还没有得到很好的研究，可能会导致目前不可预测的行为或彻底的失败，通常以车辆之间的碰撞或过载带宽消耗的形式出现。分散的机器人协调有一系列要求，这对传统的无线网络系统来说是沉重的负担：网状的连接，低延迟和高消息传输速率。这在多大程度上是一个限制因素，这是通过本研究得出的一个发现，然而，目前的无线网络标准不能满足这种需求，重点是使用协调代理来分配通话时间的高带宽集中式网络。目前没有一个主要的无线标准能够可靠地提供分散通信，在单个频带中，节点附近的消息传递速率超过每秒4000个，而在高节点数的情况下，大多数标准都能有效地达到每秒不到1000个。我的研究旨在通过研究网络和机器人控制对彼此的影响来填补这一领域的文献空白，然后提出旨在最大化两者性能的新策略。在这种情况下，争用损失可能是主要的限制，因此有许多解决方案，但是增加对机器人需求的了解开辟了选择，例如基于特征通信模式的TDMA排序。该项目的最终目标是生产可在野外大规模部署的无线协议，专门用于机器人用户。这项研究的一个必要组成部分是对机器人通信应该尊重的限制进行表征，比如信息传输速率和广播功率。由于许多可能被证明对研究至关重要的问题往往会大规模发生，因此最初的研究工作将涉及设计一个仿真，该仿真可用于在运行现有最先进的机器人控制算法的同时测试新的网络策略。模拟还必须在无线电干扰、信道衰落和反射以及其他无线限制方面包括足够的真实感。模拟到真实的组件是一个有用的工具，其中来自真实硬件的反馈用于测试环境配置的无线特性的自动表征，例如信号强度以及争用和接收概率。这将进一步用于产生机器学习模型，该模型将能够从自动确定的环境数据（平面图，天花板高度）中估计无线特性，理想情况下，这将比现有的模拟需要更少的计算资源。就未来的工业应用而言，通过该项目获得的见解应该允许调整网络协议和机器人控制算法的设计，以最大限度地提高机器人对机器人协调任务的性能和故障的可预测性。最直接的应用可能是提高道路车辆安全可靠地协调的能力。其他应用包括使用机器人履行的仓库，其中分散的网络将减少部署机器人完成任务的开销；此外，还有其他正在开发的机器人群技术，用于解决搜索和救援等无法依赖基础设施的任务。

Driver assistance technology and fully self driving vehicles are being currently being realised, and automated drone swarms are being looked to to solve complex tasks from search and rescue to cargo delivery. These technologies are being implemented largely as stand alone platforms within singular vehicles, a strategy that is likely to lead to sub-optimal behaviour as large scale deployments begin. The possibility of cooperation between vehicles using a shared data network is promising; this involves interactions between networking protocols and multi robot control algorithms. These interactions are not well studied, and may lead to currently unpredictable behaviour or outright failure, usually in the form of a collision between vehicles or overloading bandwidth consumption.Decentralised robotic coordination has a set of requirements that are taxing upon traditional wireless networking systems: mesh-like connectivity, low latency and high messaging rates. The degree to which this is a limiting factor is one discovery to be made through this research, however current wireless network standards cannot meet this constellation of requirements, with the focus being on high-bandwidth centralised networks that use a coordinating agent to assign airtime. No major wireless standard today is able to reliably provide decentralised communications where the messaging rate near a node exceeds 4000 per second in a single band, with most achieving effectively less than 1000 per second with high node counts.My research aims to fill the literature gap in this area by studying the impact that networking and robotic control have upon each other, and then propose new strategies designed to maximise the performance of both. Contention losses are likely the primary limitation in this case, for which there are many solutions, however added knowledge of robotic requirements opens up options, such as TDMA sequencing based on characterised communication patterns. The final goal of the project would be the production of wireless protocols deployable in the wild at scale specifically for robotic users. A nessecary component of the research is the characterization of the limitations that robotic communications should respect, such as messaging rates and broadcast power.Since many of the issues that are likely to prove critical for study tend to occur at scale, the initial research work will involve the design of a simulation that can be used to test new networking strategies while running existing state-of-the-art robotic control algorithms. The simulation will also have to include sufficient realism in the from of radio interference, channel fading and reflections, and other wireless limitations. A sim-to-real component is a useful tool for this, where feedback from real hardware is used for the automated characterization of the wireless properties of a test environmental configuration, such as signal strength as well as contention and reception probability. This would further be used to produce machine learnt models that would be capable of estimating wireless properties from automatically determined environmental data (floor plans, ceiling height), which would ideally require fewer computational resources than existing simulations.In terms of future application to industry, the insights gained through this project should allow the design of networking protocols and robot control algorithms to be adjusted to maximise the performance, and predictability of failures, in robot-to-robot coordination tasks. The most immediate application is likely to be in the increased ability of road vehicles to coordinate safely and reliably. Other applications include warehouses using robotic fulfillment, where a decentralised network will reduce the overhead of deploying robots for the task; in addition to other robotic swarming techniques that are in development for solving tasks like search and rescue, where infrastructure cannot be depended upon.