Probabilistic Inference Based Utility Evaluation and Path Generation for Active Autonomous Exploration of USVs in Unknown Confined Marine Environments

无人水面舰艇(USV)是能够在水面上作业的水上船只，没有任何船上的人类操作员。无人潜航器可在有限区域(港口、港口、码头等)作业。执行要求苛刻和具有挑战性的任务，如港口疏浚调查、泊位清理监测和海洋基础设施维护，其显著好处包括降低人员风险、提高时空效率和增加业务耐力。然而，当在受限的海洋环境中工作时，目前的USV通常是远程控制的。这是因为与在公共水道航行相比，受限的海洋环境是高度动态的(停靠/系泊的船只在港口的位置可能不断变化)，这使得静态海图或卫星和航空图像对航行的用处较小。这一因素使受限的海洋环境变得更加不为人所知，并与高度的不确定性有关。自主探测作为一种可以自动绘制未知受限环境地图的过程，因此成为在未知受限海洋环境中进行USV操作的关键。目前USV采用的最先进的自主探索策略是利用同步定位和测绘(SLAM)技术，在没有任何先验信息的情况下使用感官数据构建环境地图。由于SLAM是一个被动的过程，现有的USV平台需要定期的遥操作，由操作员指导地图构建过程，使得探索不是完全自主的。为了使基于SLAM的自主探索成为主动过程，必须集成包括两个模块的规划功能，即效用评估模块和路径生成/选择模块。然而，目前有关效用评估和路径生成的研究不能解决海洋环境中地标稀疏带来的问题，这将影响勘探的精度和效率。因此，本研究旨在利用基于概率推理的效用评估和路径生成/选择来开发一种新的主动自主探索框架。更具体地说，我们将构建一个包含虚拟地标的伪地图，作为具有稀疏真实地标的未知受限海洋环境的代理，并使用贝叶斯概率推理分别根据虚拟地标的姿态和位置的边际后验分布来评估不确定性。我们还提出了一种新的基于高斯过程(GP)的自主探索路径生成算法，并将路径生成问题作为因子图上的概率推理来解决。交叉熵优化方法将适用于路径规划，通过考虑诸如USV的操纵性等非线性约束，使得能够有效地推导GP均值和协方差更新规则。这项工作成功的关键是与史蒂文斯理工学院领先的海洋机器人专家Brendan Englot教授的国际合作，共同开发该框架。这项工作还将与经验丰富的行业合作伙伴密切合作，包括伦敦港务局(Port Of London Authority)和BMT Group Ltd。通过与解放军和BMT密切合作，这项研究产生的创新将在Otter USV上实施，以进行泰晤士河潮汐测量的用例演示(如水文测量)。这一国际合作的长期愿景是建立一个强大的英美研究联盟，研究未来先进传感器、人工智能/机器学习和机器人领域的海洋创新，与更多的学术机构、公司和监管机构/组织合作。

Unmanned surface vessels (USVs) are water-borne vessels that are capable of operating on the surface of the water without any onboard human operators. USVs can operate in confined areas (ports, harbours, marinas, etc.) to conduct demanding and challenging missions such as port dredging survey, berth clearance monitoring and marine infrastructure maintenance, with significant benefits including reduced risk to personnel, improved spatial-temporal efficiency and increased operation endurance. However, when operating in confined marine environments, current USVs are usually remotely controlled. This is because in contrast to navigating in public waterways, confined marine environments are highly dynamic (the locations of docked/moored vessels in a port may be constantly changing) making static nautical charts or satellite and aerial imagery less useful for navigation. Such a factor makes the confined marine environment more highly unknown and associated with high levels of uncertainties. Autonomous exploration, as a process that can map an unknown confined environment in an automatic way, has therefore become critical to USV operation in unknown confined marine environments. Current state-of-the-art autonomous exploration strategy employed by USVs is to leverage the Simultaneous Localisation And Mapping (SLAM) technology to build a map of an environment using sensory data without any prior information. Since SLAM is a passive process, regular teleoperation with human operators guiding the map-building process is required for existing USV platforms, making the exploration not fully autonomous. To make the SLAM based autonomous exploration an active process, planning functionality including two modules, i.e., a utility evaluation module and a path generation/selection module, has to be integrated. However, current studies about utility evaluation and path generation cannot address the issues caused by the sparse landmarks in a marine environment, which will compromise the exploration accuracy and efficiency. This research therefore aims to develop a new active autonomous exploration framework using probabilistic inference based utility evaluation and path generation/selection. More specifically, we will construct a pseudo map which contains virtual landmarks as a proxy for an unknown confined marine environment with sparse real landmarks, and evaluate uncertainties as per marginal posterior distributions of poses and positions of virtual landmarks, respectively, using Bayesian probabilistic inference. We also propose to design a new Gaussian Process (GP) based path generation algorithm for autonomous exploration and solve the path generation problem as probabilistic inference on a factor graph. A cross-entropy optimisation method will be adapted to the path planning to enable efficient derivation of the GP mean and covariance updating rules by taking into account nonlinear constraints such as USVs' manoeuvrability. Of key importance for the success of this work is the international collaboration with a leading marine robotics expert, Prof. Brendan Englot, Stevens Institute of Technology, to jointly develop the framework. This work will also have a close collaboration with experienced industrial partners, including Port of London Authority (PLA) and BMT Group Ltd. By working closely with PLA and BMT, innovations generated from this research will be implemented on the Otter USV to conduct use-case demonstrations (e.g., hydrographic survey) on the Tidal Thames. And the long-term vision of this international collaboration is to establish a strong UK-US research consortium on future marine innovations in advanced sensors, AI/machine learning and robotics to work collaboratively with more academic institutions, companies and regulators/organisations.