Related Work

Several categories of work relate to orbital edge computing. Section 2 provides a brief space systems background. The NASA guide to the state-of-the-art in small satellites [9] describes characteristic aspects of nanosatellites, including technology readiness levels of essential subsystems. Recent surveys [5, 62] study multi-satellite orbital dynamics and provide an overview of propulsion systems. Commercial efforts demonstrate the viability of camera-equipped nanosatellite constellations [20, 74]. The SpaceNet challenge [94] illustrates broad interest in visual computing on space data, and the proposed Amazon ground station network [2] further cements the value of computing on visual and other space sensor data.

Recent edge computing work provides context for our work. Edge computing recognizes that, as high-datarate sensors (e.g. cameras, lidar) proliferate, streaming all data to central cloud systems for processing becomes infeasible [78, 79, 97]. Edge computing is important particularly in cases of complex processing, e.g. video querying [44], search [57], or DNN speech processing [55]. We propose to leverage early discard, which has been studied for search [45], video indexing [44], and drone video processing [97]. Recent work [8, 82] demonstrates the utility of simulation frameworks for edge computing on drones; cote is an analogous utility for the orbital edge. Machine inference accelerators [14, 15, 24, 33] could significantly shorten full-coverage CNP pipeline depths, although some that rely on temporal data redundancy [10] may have limited benefit for devices capturing images at the GTFR.

Intermittent computing [61] shares challenges with orbital edge computing in that both types of systems are fully energy-autonomous. A number of recent intermittent systems [16, 39, 54, 64–66] function despite unpredictable power failures using techniques that may be applicable to the orbital edge in future work. Some intermittent computing platforms [17] are similar in that they rely only on supercapacitors for energy storage. Other intermittent systems are similar in that they target DNN workloads [32] and communication minimization [63] at the edge in batteryless devices. While similar in spirit, these efforts differ significantly in scale, deployment environment, and in their inability to rely on processor sleep modes; instead, they power off frequently.

与轨道边缘计算相关的研究可分为几类。第2节提供了简要的空间系统背景。NASA的小型卫星技术发展水平指南[9]描述了纳米卫星的典型特征，包括关键子系统的技术成熟度等级。近期的综述[5, 62]研究了多卫星轨道动力学，并概述了推进系统。商业领域的实践证明了配备相机的纳米卫星星座的可行性[20, 74]。SpaceNet挑战赛[94]表明了学界对空间数据视觉计算的广泛兴趣，而亚马逊提议的地面站网络[2]则进一步巩固了对视觉及其他空间传感器数据进行计算的价值。

近期的边缘计算研究为我们的工作提供了背景。边缘计算领域认识到，随着高数据率传感器（例如相机、激光雷达）的激增，将所有数据流式传输到中心云系统进行处理变得不可行[78, 79, 97]。 边缘计算在处理复杂任务时尤为重要，例如视频查询[44]、搜索[57] å或DNN语音处理[55]。 我们提议利用“早期丢弃”（early discard）策略，该策略已在搜索[45]、视频索引[44]和无人机视频处理[97]等领域得到研究 。近期的工作[8, 82]展示了仿真框架在无人机边缘计算中的效用；cote则是轨道边缘场景下的一个类似工具。机器推断加速器[14, 15, 24, 33]可以显著缩短实现完全覆盖的CNP流水线深度，尽管一些依赖时间数据冗余[10]的加速器对于以GTFR速率捕获图像的设备来说，其效益可能有限。

间歇式计算（Intermittent computing）[61]与轨道边缘计算有共通的挑战，因为这两类系统都是完全能量自主的。许多近期的间歇式系统[16, 39, 54, 64–66]尽管会面临不可预测的电源故障，但仍能正常工作，其使用的技术可能在未来的工作中适用于轨道边缘。一些间歇式计算平台[17]与我们的系统相似，它们同样仅依赖超级电容器进行储能。其他间歇式系统则在针对无电池设备上的DNN工作负载[32]和最小化通信[63]方面具有相似性。尽管理念相似，但这些工作在规模、部署环境以及无法依赖处理器睡眠模式（而是频繁断电）方面存在显著差异。