AISat: enabling spacecraft autonomy with smart perception

从2004年到2012年，平均每年大约有72个物体被送入近地轨道。然而，自2012年以来，放置在近地轨道上的卫星数量急剧增加，现在每年的卫星数量达到125颗，为了实现快速互联网，几家公司正在近地轨道上部署由1000多颗卫星组成的大型星座。限制发射速率是不可行的，因为它不能被强制执行。因此，为了减少LEO中大型物体的数量，唯一的选择是主动移除目前在轨的大型物体。为了实现这一目标，需要发展技术，包括先进的图像处理、复杂的制导、导航和控制以及捕捉碎片的创新机器人。由萨里大学高阳教授创立的AI4SPACE公司正在开发AISat 3D视觉感知和导航系统。这可以使用卫星内部已经存在的空间相机，因此对于清除碎片以及空间回收、航天器加油和卫星健康检查来说，是一种更具成本效益、可持续的软件定义解决方案。该项目将进一步开发我们的尖端导航软件，用于地球轨道小卫星任务（例如CubeSat）的商用现货（COTS）空间合格硬件，并与在轨服务（IOS）任务运营商和航天器平台制造商形成潜在的合作伙伴关系。我们的AISat技术通过应用所谓的“深度学习”方法推动了经典视觉导航方法的边界，AISat软件框架利用简化的图像数据处理管道，包括轨道特征检测，关键点估计和卷积神经网络来估计空间目标的方向。萨里大学已经使用TANGO和联盟号宇宙飞船的真实数据验证了一个研究原型。我们的解决方案在esa -斯坦福“大挑战”中被评为世界上最好的解决方案之一，获得了前三名的位置。我们已经确定了2025-2027年间的86个IOS任务概念和设计将需要3D空间感知，我们将利用该项目的资金来整合和验证软件，为2023- 2024年的在轨演示机会做好准备。

Averaged over the eight years 2004--12, about 72 objects were placed into LEO (Low Earth Orbit) per year. However, since 2012, there has been a steep increase in the number of satellites placed in LEO, with the count now running at 125 objects per year and for fast internet several companies are deploying large constellations of more than 1000 satellites in LEO. Limiting the launch rate is not feasible because it cannot be mandated. Therefore, in order to reduce the number of big objects in LEO, the only option is to actively remove large objects now in orbit. To achieve this, technology developments are required, including advanced image processing, complex guidance, navigation and control and innovative robotics to capture debris.AI4SPACE, a spinout founded by Professor Yang Gao from University of Surrey, is developing the AISat 3D visual perception and navigation system. This can use the space camera already in place within satellites and hence represents a more cost-effective, sustainable software-defined solution for debris removal as well as for space recycling, spacecraft refuelling and satellite health checks.This project will further develop our cutting edge navigation software to be used on Commercial off The Shelf (COTS) space qualified hardware for Earth orbital small satellite missions (e.g. CubeSat) and form potential partnerships with In-orbit servicing (IOS) mission operators and spacecraft platform manufacturers. Our AISat technology pushes the boundary of classical visual navigation approach by applying so called "deep learning" methods and the AISat software framework utilises a streamlined image data processing pipeline including orbital feature detection, keypoint estimation and convolutional neural networks to estimate the orientation of the space target. University of Surrey already has a research prototype validated using real-world data from the TANGO and Soyuz spacecraft. Our solution has been ranked one of the best solutions in the world within the ESA-Stanford "Grand Challenge" taking a top three position.We have identified that 86 IOS mission concepts and designs from 2025-2027 will need 3D space perception and we will use the funding from this project to integrate and validate the software ready for in-orbit demonstration opportunities in a 2023-24 timeframe.