Towards More Autonomy for Unmanned Vehicles: Situational Awareness and Decision Making under Uncertainty

预计在不久的将来，无人驾驶车辆将广泛应用于军事和民用行动，并对我们的日常生活产生深远的影响。在充分发挥无人驾驶汽车带来的潜力之前，我们可以合理地预期，要让无人驾驶汽车被用户、公众和监管部门所接受，它们必须达到与人类操作系统相似的安全水平。在许多其他方面，无人驾驶车辆的基本要求是以安全，及时和适当的方式响应内部和外部变化的能力。因此，态势感知和决策是无人驾驶车辆安全运行的两项最重要的使能技术。在很大程度上，它们决定了无人驾驶汽车的自主和智能水平。与居住在车辆中的人类驾驶员或飞行员相比，一个主要的安全问题是位于远程控制站的无人驾驶车辆操作员的态势感知能力不可避免地降低。无人驾驶车辆在动态的、不可预测的环境中运行，其感官信息不完整（或不准确），这给态势感知和决策带来了许多挑战。概率法和有界法分别被广泛用于表示具有已知分布或已知上界和下界的不确定性。态势感知包括在一定时间和空间内对环境中物体的感知，对其意义的理解以及对其近期状态的预测。例如，在对感兴趣的移动物体的近距离状态的投影中，任何与感知和理解相关的初始不确定性将随着投影时间跨度的增加而呈指数级扩展。然而，通过利用世界模型中的信息，如操作环境、道路规则（或空中规则）和已识别对象的属性，有可能显著减少不确定性。对于概率不确定性，这使得高斯分布假设无效，这是目前大多数统计方法（如卡尔曼滤波）的基础。在高斯分布假设下，运动物体的估计状态可以用其均值带方差来表示，并且可以定义一个对称的不确定区域，该区域的均值位于中心（在指定的置信水平下，如99%）。知识的引入（例如，由于道路布局的约束）使这种情况不再成立。为了解决利用世界模型信息带来的非高斯分布的挑战，将开发一个严格的贝叶斯学习框架，用于汇集世界模型和测量数据的所有知识，以提供对环境的更好估计，并随着投影时间传播不确定区域。可达性分析将针对最坏情况分析的有限不确定性进行开发，其中不确定性将使用来自世界模型的约束来减少。将进行危害分析，以识别任何潜在的风险。关键思想是采取积极主动的方法，通过提高态势感知，推理和决策来防止任何紧急情况。态势感知提供的估计和相关的不确定区域将被馈送到新的决策和规划工具中。研究活动将得到小型地面和空中交通工具试验的大力支持和验证。该项目旨在显著提高无人驾驶车辆运行的安全水平，弥合无人驾驶车辆在现实应用中的开发和部署之间的差距，这是新的业务增长的战略重要领域。

It is anticipated that unmanned vehicles will be widely used within military and civilian operations and have a profound influence in our daily life in near future. Before fully realising the potential that unmanned vehicles bring, it is reasonably expected that to make unmanned vehicles accepted by users, the public and regulatory authorities, they shall achieve a similar level of safety as human operated systems. Among many others, a fundamental requirement for an unmanned vehicle is the capability to respond to internal and external changes in a safe, timely and appropriate manner. Therefore, situational awareness and decision making are two of the most important enabling technologies for safe operation of unmanned vehicles. To a large extent, they determine the level of autonomy and intelligence of an unmanned vehicle. Compared with a human driver or pilot residing in the vehicle, a major safety concern is the inevitable reduction in situational awareness of the unmanned vehicle operator remotely located in a control station. Unmanned vehicles operate in a dynamic, unpredictable environment with incomplete (or inaccurate) sensory information, which creates many challenges in situational awareness and decision making. Probabilistic and bounded approaches are widely used to represent uncertainty with a known distribution or with a known upper and lower bounds respectively. Situational awareness includes the perception of the objects in the environment within a volume of time and space, the comprehension of their meaning and the projection of their status in the near future. For example, in projection of the near status of moving objects of interest, any initial uncertainty associated with perception and comprehension will expand exponentially with the increase of the projection time span. However, it is possible to significantly reduce the uncertainty by utilising the information in the world model such as the operation environment, the Rules of the Road (or of the Air) and the properties of an identified object. For probabilistic uncertainty, this makes the Gaussian distribution assumption invalid, which is fundamental for most of the current statistical approaches such as Kalman filtering. Under the Gaussian distribution assumption, the estimated state about a moving object can be presented by its mean with a variance, and a symmetric uncertain region can be defined with the mean located at the centre (under a specified confidence level such as 99%). The introduction of knowledge (e.g. constraints due to the roadway layout) makes this not true anymore. To address the challenge of non-Gaussian distributions imposed by making use of information from the world model, a rigorous Bayesian learning framework will be developed for pooling all the knowledge from the world model and measurement data to provide a better estimate of the environment, and to propagate the uncertain regions with projection time. Reachability analysis will be developed for bounded uncertainty for worst case analysis, where the uncertainty will be reduced using constraints from the world model. Hazard analysis will be carried out to identify any potential risk. The key idea is to take a proactive approach to prevent any emergent situation through improving situational awareness reasoning and decision making. The estimates and associated uncertain region provided by the situational awareness will be fed to novel decision making and planning tools. The research activities will be strongly supported and verified by experimental tests on small scale ground and aerial vehicles. This project aims to significantly improve the level of safety of unmanned vehicle operation and to bridge the gap between the development and deployment of unmanned vehicles in real world applications, which is a strategically important area for new business growth.