Consider large groups of unmanned underwater vehicles (UUVs) conducting large-scale geospatial tasks, such as information gathering or area sensing. Major costs of long duration missions include expensive underwater positioning systems and propulsion, which consumes energy. Exploiting the ocean currents can increase endurance, but requires accounting for forecast uncertainty, which lies beyond the scope of this article. State-of-the-art techniques, such as Monte Carlo tree search or cross entropy method that coordinate underwater vehicles for path-dependent rewards, do not scale well to such large groups. Furthermore, solving the mentioned tasks requires accounting for overlaps in the areas each vehicle searches, increasing the complexity of the problem. We therefore investigate planning techniques that can evaluate path-dependent rewards, account for the ocean forecast, and efficiently coordinate plans for many agents. Two formulations are investigated, which either search the space of action sequences or the space of feedback policies to find dynamically feasible trajectories. We present what we believe to be the first application of the cross entropy method to create joint plans for large groups of 8-128 UUVs. We also develop a novel iterative greedy method that further refines the best discovered constant action sequences to improve other greedy techniques. The iterative greedy method gathers the most information on average, scales well to deploying large groups of agents, and gathers 3%-8% more reward than other techniques.

中文翻译：

---

协作水下航行器执行大规模地理空间任务


考虑大型无人水下航行器 (UUV) 执行大规模地理空间任务，例如信息收集或区域感测。长时间任务的主要成本包括昂贵的水下定位系统和推进装置，这会消耗能量。利用洋流可以提高耐力，但需要考虑预测的不确定性，这超出了本文的范围。最先进的技术，例如协调水下航行器以获得路径依赖奖励的蒙特卡罗树搜索或交叉熵方法，不能很好地扩展到如此大的群体。此外，解决上述任务需要考虑每辆车搜索区域的重叠，这增加了问题的复杂性。因此，我们研究了可以评估路径相关奖励、考虑海洋预测并有效协调许多代理的计划的规划技术。研究了两种公式，它们要么搜索动作序列的空间，要么搜索反馈策略的空间，以找到动态可行的轨迹。我们提出了交叉熵方法的首次应用，用于为 8-128 个 UUV 的大型群体创建联合计划。我们还开发了一种新颖的迭代贪婪方法，该方法进一步细化最佳发现的恒定动作序列，以改进其他贪婪技术。迭代贪婪方法平均收集最多的信息，可以很好地扩展到部署大型代理组，并且比其他技术多收集 3%-8% 的奖励。