Decentralized Cooperative Exploration of Nonstationary Spatiotemporal Environmental Fields with Autonomous Underwater Vehicle Systems

用于环境场估计的系统非常有兴趣，因为它们可以在海洋研究、定位污染源和分析海上风电场的流场中收集数据。通常，安装固定的传感器来完成这类任务。然而，由于可能的传感器重新定位，一组自主水下机器人(AUV)可以获得更高的空间分辨率。虽然由中央单元控制的移动传感器网络可以获得很好的性能，并且已经得到了深入的研究，但在水下场景中，由于通信信道不可靠且带宽低，分散的方法往往受到青睐，因为机器人无法与中央单元定期通信。为了开发真正分散的多AUV系统，需要在现有技术基础上进行扩展。本研究项目考虑了一组部署在海洋环境中执行分布式时空估计任务的AUV。AUV对动态物理过程进行测量，这些过程被建模为非平稳的高斯随机场。这些场描述了水温或污染物浓度等量的演变。利用高斯随机场的协方差，基于同步和路径积分控制，实现AUV群的分散协调。一种基于同步和路径积分控制的方法可以实现分散的环境探索。由于未知场是由非参数方法(高斯过程)建模的，因此需要最少的先验知识。可以沿场估计初始未知的协方差函数。这种方法允许分析探索性问题的基本权衡，总结为以下问题：是应该收集未来的观测数据来改进场模型(更好的协方差函数估计)，还是应该在当前(可能不准确的)模型表示最高不确定性的位置进行观测？近年来，力学和海洋工程研究所进行了关于AUV动力学和控制的理论和实践方面的研究并发表了论文。推导了用于探索环境领域的集中控制的AUV群的算法，并用基于物理的方法进行了建模。微型AUV已经被开发出来。这些成果是开发一个真正分散的勘探控制框架的起点。此外，还对将现场不确定性与分散协调控制相结合进行了初步研究。进一步的研究表明，车辆可以探索平流扩散场，其中场信念由高斯马尔可夫随机场表示，这一概念将进行实验验证。在开发的控制框架下，一组微型AUV将执行水平曲线跟踪、梯度爬升和源定位等任务。

Systems for environmental field estimation are of great interest because they allow data collection in oceanographic research, locating pollution sources, and analyzing flow fields at offshore wind farms. Usually, stationary sensors are installed for such tasks. However, a group of Autonomous Underwater Vehicles (AUVs) can achieve a higher spatial resolution due to possible sensor repositioning. While mobile sensor networks controlled by a central unit can achieve great performance and are already investigated in great depth, decentralized approaches are often favored in underwater scenarios where vehicles cannot communicate with a central unit regularly, because communication channels are unreliable and of low bandwidth. The development of truly decentralized multi-AUV systems requires extensions beyond the state of the art.A group of AUVs deployed into a marine environment performing distributed spatiotemporal estimation tasks is considered in this research project. The AUVs take measurements of dynamic physical processes which are modelled as non-stationary Gaussian random fields. The fields describe the evolution of quantities such as water temperature or concentration of pollutants. The covariance of the Gaussian random field is used for decentralized coordination of the AUV group based on synchronization and path integral control. An approach based on synchronization and path integral control enables decentralized environmental exploration. Since the unknown fields are modeled by a nonparametric approach (Gaussian processes) a minimum of a priori knowledge is required. The initially unknown covariance function can be estimated along the field. The approach allows analyzing a fundamental tradeoff in explorative problems which is summarized by the following question: Shall future observations be collected to improve the field model (a better covariance function estimate), or shall observations be taken at locations where the current (possibly inaccurate) model indicates the highest uncertainty?The Institute of Mechanics and Ocean Engineering has in recent years conducted and published research on both theoretical and practical aspects of AUV dynamics and controls. Algorithms for centrally controlled AUV groups for exploring environmental fields have been derived which are modeled with physics-based approaches. Micro AUVs have been developed. These results are the starting point for developing a truly decentralized controls framework for exploration. Furthermore, preliminary studies on combining field uncertainty with decentralized coordination control have been conducted. Further studies show that a vehicle can explore an advection-diffusion field where the field belief is represented by a Gaussian Markov random field.The verification of the concept will be conducted experimentally. With the developed controls framework, a group of micro-AUVs will perform tasks such as level curve tracking, gradient climbing and source localization.