Autonomous robotics in noisy and delayed environments

该研究旨在开发能够在所有环境中开发和部署自主机器人的方法和流程，其中机器人可以有助于安全和生产力，特别是在延时环境中。时间延迟来自几个不同的来源，如传感器，计算，执行器，通信和系统不同部分之间的交互，特别是机器人到机器人（或代理到代理）的交互。我们主要感兴趣的情况下，在线实时学习应用于多机器人环境中出现的个人和群体行为的机器人集体出现，并在其中的时间延迟可能会导致不同的，甚至不稳定的行为，单独或集体。 出于本申请的目的，个体行为由单个机器人在其感知环境处于给定状态的情况下如何选择行动（在策略或策略方面）来定义。机器人感知环境状态的方式本质上是不精确的，有限的，嘈杂的和潜在的时间延迟。此外，其他机器人的存在会改变环境，以至于与另一个机器人的关系可能会影响在任何给定时间做出的决定。群体行为是当面临相互冲突的角色和任务时，集体如何作为一个群体行事。群体行为与个人行为的联系方式目前尚不完全清楚。 本研究旨在应用基于强化学习技术（及其衍生物）的学习算法，以产生在线实时能力的机器人，可应用于实际应用，如简易爆炸装置（IED）的定位和破坏，物体识别，矿井和隧道导航，空中监视，水下测绘，拥挤环境中的导航等。 环境的表示是通过使用学习算法以及基于模型预测控制（MPC）的分散控制来完成的，特别是当考虑代理之间的相互作用时（合作-当代理具有共同的目标或目的时-或竞争-当代理没有明确表示其他人的目标时-本质上）。学习将应用于两个不同的层：（i）任务的单独执行;以及（ii）代理之间的交互。 开发的算法将在符合机器人操作系统（ROS）的真实的机器人平台上进行测试。这些算法需要在有噪声和延迟的传感器读数的情况下实时运行。此外，在控制算法的处理中的随机延迟是可能的。该平台由地面和空中车辆组成，实验设施已经在加拿大皇家军事学院（RMCC）提供。

This research aims to develop the methods and processes that enable the development and deployment of autonomous robots in all environments wherein robots can contribute to safety and productivity, especially in time-delayed environments. The time delay comes from several different sources as sensors, computation, actuators, communication and interactions among different parts of the system, especially robot to robot (or agent to agent) interaction. We are primarily interested in the case of on-line real-time learning applied to multi-robot environments in which emergence of individual and group behaviours in collectives of robots arise and in which time-delays could lead to different, or even unstable, behaviours, individually or collectively. For the purposes of this application, individual behaviour is defined by how a single robot chooses to act (in terms of policies or strategies) given that it perceives the environment to be at a given state. The way the robot perceives the state of the environment is inherently imprecise, limited, noisy and potentially time-delayed. Also, the presence of other robots change the environment in such a way that the relationship with another single robot may affect the decision taken at any given time. Group behaviour is how the collective may act as a group when faced to conflicting roles and tasks. Group behaviours are linked to individual behaviours in a manner that is not presently completely known. This research intends to apply learning algorithms based on reinforcement learning techniques (and its derivatives) in order to produce on-line real-time capable robots that can be applied to actual applications such as the location and disruption of improvised explosive devices (IED), object identification, mine and tunnel navigation, aerial surveillance, underwater mapping, navigation in crowded environments and so on. The representation of the environment is to be done by using learning algorithms as well as decentralized control based on Model Predictive Control (MPC), especially when considering the interactions among agents (either cooperative - when the agents have a common goal or objective - or competitive - when the agents do not have an explicit representation of the goal of the others - in nature). The learning will be applied to two different layers: (i) the individual execution of a task; and (ii) the interaction among agents. The algorithms developed will be tested in real robotic platforms that are compliant with the Robot Operating System (ROS). These algorithms are needed to run in real-time with the noisy and delayed sensor readings. Furthermore, stochastic delays in the processing of the control algorithms are possible. The platforms are composed of ground and aerial vehicles and the experimental facility is already available at the Royal Military College of Canada (RMCC).