NRI: Information-Theoretic Trajectory Optimization for Motion Planning and Control with Applications to Space Proximity Operations

无论是在地球轨道上还是在地球轨道以外的许多任务中，太空机器人操作都是不可或缺的。卫星维修和加油、空间站补给消耗品、清除空间碎片、航天器结构完整性检查、机组人员协助以及对火星和其他行星和彗星的深空任务的支持，所有这些都需要高精度、可靠和自主(或半自动)的空间机器人的协助。到目前为止，大多数太空中的机器人操作都是由操作员在严密监督的模式下进行的。这限制了可以执行的任务的灵活性和类型(例如，光往返火星需要大约15分钟，因此不可能进行实时的远程控制)。这项研究的目的是发展必要的理论和算法，以便能够通过对自由飞行的空间机器人在另一物体附近进行稳健、可靠的传感和规划来利用主动探索，以便执行近距离作业(包括在空间中的自主交会和对接)。这类问题的挑战之一是，为了计划适当的控制行动，在了解周围环境方面存在不确定性。为了应对这些挑战，我们利用了随机最优控制领域的新工具和方法，以及描述航天器姿态动力学和在轨航天器运动学的新进展。为了确保我们开发的算法在现实生活中如预期的那样执行，理论结果将在高保真的5自由度航天器模拟器上进行实验验证。这项工作将通过推进在另一个人造或天然物体附近轨道航天器的感知和路径规划方面的最先进技术，对美国在太空中进行常规卫星监测和维修的能力产生直接影响。虽然这项工作的重点主要是空间机器人的应用，但同样的技术也可以用于所有类似的问题，即智能代理需要在不确定和动态的环境中自主导航。目前文献中的范例基本上利用知觉线索(特别是那些仅基于视觉信息的)作为全状态反馈估计器的替代品，从而强制将知觉和控制行为人为地分离。这种感觉数据采集/处理和控制/驱动策略之间的二分法-从它对受加性噪声(分离原理)约束的线性系统的稳定的广泛适用性-深深植根于社区，不适合于这个问题，在这个问题中，信息收集(感知/传感)与运动(控制)紧密耦合。为了克服上述局限性，在本工作中，建议使用随机最优控制的工具来从原始感觉输入中提取可操作的信息。该方法的一个关键是跟踪估计误差的一阶和二阶统计量，并将它们视为状态，从而使控制行动依赖于这两个统计量。其结果是一种新的、在计算上更有效的方法，以在探索阶段最大限度地收集信息，并在执行阶段优化轨迹的分布。

Robotic operations in space are indispensable for many missions both in Earth orbit and beyond. Satellite servicing and refueling, space station resupply with consumables, removal of space debris, spacecraft structural integrity inspection, crew assistance, as well as support for deep space missions to Mars and other planets and comets, all require the assistance of highly accurate, reliable and autonomous (or semi-autonomous) space robots. To date, most robotic operations in space are performed in a closely supervised mode by a human operator. This limits both the flexibility and the type of missions that can be performed (for example, the time for light to travel to and from Mars takes about 15 minutes, making "real-time" remote control impossible). This research aims at developing the necessary theory and algorithms to be able to utilize active exploration using robust, reliable sensing and planning of a free-flying space robots in the vicinity of another body, in order to perform proximity operations (including autonomous rendezvous and docking in space). One of the challenges in these types of problems is the uncertainty in understanding the surroundings in order to plan suitable control actions. In order to handle these challenges we utilize novel tools and methodologies from the field of stochastic optimal control along with new advances describing the spacecraft attitude dynamics and kinematics of spacecraft in orbit. In order to ensure that the algorithms we develop perform in real-life as expected, the theoretical results will be experimentally validated on a high-fidelity 5-dof spacecraft simulator facility. This work will have an immediate impact on the US capabilities to perform monitoring and servicing of satellites in space routinely, by advancing the state-of-the-art in perception and path-planning of orbiting spacecraft in the vicinity of another body, man-made or natural. Although the emphasis of this work is primarily on space robotic applications, the same techniques can be used in all similar problems where an intelligent agent needs to navigate autonomously in an uncertain and dynamic environment.The proposed research tackles a fundamental problem in autonomous/robotic systems, namely, the integrated sensing and planning under uncertainty. The current paradigm in the literature utilizes perceptual cues (especially those based solely on visual information) essentially as surrogates of full-state feedback estimators, thus enforcing an artificial separation of perception and control action. This dichotomy between sensory data acquisition/processing, and control/actuation strategies - deeply rooted in the community from its wide applicability to the stabilization of linear systems subject to additive noise (?separation principle?) - is unsuitable for this problem, where information gathering (perception/sensing) is tightly coupled with motion (control). To overcome the aforementioned limitations, in this work it is proposed to use tools from stochastic optimal control in order to extract actionable information from raw sensory inputs. A key ingredient of the proposed approach is to keep track of the first and second order statistics of the estimation error and treat them as the state, so that control actions depend on both of them. The result is a new, computationally more efficient, methodology to maximize information gathering during the exploration phase and to optimize over distributions of trajectories during the execution phase.