RI: Small: Decision Making with Spatially and Temporally Uncertain Data

这项研究探讨了机器人如何在动态环境中有效地进行规划。重点是利用自主水下机器人进行海洋时空监测。这样的监测和传感使科学家能够更好地了解地球及其环境过程。为了在海洋中自主操作，机器人需要做出决定，尽管洋流在空间和时间上都有很大的不确定性，而且通常足以显著改变车辆的运动。在这项工作中开发的算法在水下领域与感兴趣的海洋学现象（例如，藻华）进行了测试。测试是在南加州大学水下机器人先前现场试验的多个数据集上进行的。通过使用机器人测试和模拟测试，在这项工作中开发的方法在广泛的应用和规模上得到了验证。在规划一项长期的环境监测任务时，准确的规划者需要对目的和中间航路点的洋流进行预测，以便确定相应的行动。最先进的决策制定方法的一个缺点在于，它们依赖于已知的不确定性描述，即“某个时刻的快照”。在这项研究中，重点是设计新的机制，推广最先进的决策方法，以克服这一限制。目标是创建一种通用的方法来计算控制策略，不仅考虑完全已知的（当前和过去）随机描述，而且考虑（可能不确定的）未来动态预测。随机过渡动力学被认为是一个随时间等参数变化的有噪声或不确定的函数。为了解决这一问题，本文提出了一种新的、高效的价值传播机制，涉及空间和时间两个维度的演化过程。

This research investigates how robots can plan effectively in dynamic settings. The focus is on spatio-temporal ocean monitoring with autonomous underwater robots. Such monitoring and sensing allows scientists to gain a greater understanding of the planet and its environmental processes. For autonomous operation in the ocean, robots need to make decisions in spite of extensive uncertainty in ocean currents that change both spatially and temporally, and are often strong enough to alter the vehicle's motion significantly. The algorithms developed in this work are tested in the underwater domain with oceanographic phenomena of interest (e.g., algal blooms). Tests are performed on multiple datasets available at USC from prior field trials with underwater robots. Through the use of both robot testing and simulated testing, the methods developed in this work are validated across a wide range of applications and scales.When planning for a long-range and long-term environmental monitoring task, an accurate planner requires a forecast of ocean currents at the destination and intermediate way-points in order to determine corresponding actions. A drawback of state-of-the-art decision making methodologies lies in the fact that they rely on a known uncertainty description that is a "snapshot at a certain moment in time." In this research, the focus is on the design of new mechanisms that generalize state-of-the-art decision-making methodologies to overcome this limitation. The goal is the creation of a general methodology to compute control policies that consider not only a fully known (current and past) stochastic description, but also a (possibly uncertain) prediction of future dynamics. The stochastic transition dynamics are considered as a noisy or uncertain function that varies with some parameter such as the time. To solve this problem, a new and efficient value propagation mechanism is developed involving two processes that evolve in both spatial and temporal dimensions.