基于视觉的空间非合作目标的智能捕获技术研究

On-orbit service and maintenance of spacecraft are an important research issue in aerospace field, and are valuable to both military and civilian applications. Capture of space object is the premise of on-orbit service and maintenance. In many on-orbit service missions, such as on-orbit assembling, breakdown maintaining, refueling, module replacing, technology upgrading, and auxiliary mechanism deploying, space target should be captured at the beginning of the mission. Since non-cooperative space target (such as space trash, disable and failed satellites) cannot communicate with and cooperate with the service spacecraft, the capture of them is a very challenging mission. To meet the significant demand of Chinese aerospace industry in the future, this project studies the intelligent capture technology of non-cooperative space target. In this research, the computer vision technology is employed to study the 3D reconstruction of the target, the optimization method is used to study the motion prediction of the target based on vision information, the deep learning and reinforcement learning technologies are adopted to study the intelligent grasping of the target, and the experiment is conducted to validate the proposed 3D reconstruction method. The research of this project can lay a foundation for on-orbit service and maintenance, and promote the development of Chinese aerospace engineering.

航天器的在轨服务与维护是航天领域未来重要的发展方向，具有极高的军民两用双重价值。空间目标的在轨捕获是完成航天器在轨服务与维护的前提，如对卫星进行在轨装配、故障维修、燃料加注、模块更换、技术升级以及辅助机构展开等，都需要首先完成空间目标的在轨捕获，否则在轨服务无从谈起。由于空间非合作目标（如空间垃圾、失效航天器、失控航天器）在信息层面上不沟通、在机动行为上不配合，对其的捕获更具挑战性。本申请项目面向我国航天未来发展的重大需求，对空间非合作目标的智能捕获技术进行研究，采用计算机视觉技术研究非合作目标的三维结构重建方法，基于视觉信息和采用优化方法研究非合作目标的运动预测方法，采用深度学习和强化学习两种人工智能技术研究非合作目标的智能捕获方法，并且开展相关方法的实验验证，为实现航天器在轨服务与维护奠定理论基础，寄望于研究成果能够促进我国航天事业的发展。

航天器的在轨服务与维护是航天领域未来重要的发展方向，具有极高的军民两用双重价值。空间目标的在轨捕获是完成航天器在轨服务与维护的前提，如对卫星进行在轨装配、故障维修、燃料加注、模块更换、技术升级以及辅助机构展开等，都需要首先完成空间目标的在轨捕获，否则在轨服务无从谈起。本项目针对空间机器人在轨非合作目标任务中，抓捕前阶段的运动观测与预测、抓捕中阶段的抓捕结构智能识别与操作控制问题进行了深入研究并取得了丰富的研究成果。具体成果如下：1.提出了一种改进最近点迭代算法来解决表面覆有高反射率材料的非合作目标位姿观测问题；2.提出了一种高精度两步的非合作目标的运动预测方法；3. 给出了一种快速生成卫星抓捕结构识别训练用数据的构建方法，并基于所构建数据对多种结构识别神经网络的识别精度进行了有效评估，评估结果为工程部门解决抓捕结构的智能识别算法的选择提供参考；4.提出两种可实现接触保持的碰撞控制方法，经验证两种方法能确保空间机器人顺利完成抓捕操作。上述研究成果均是以实际需求为出发点，具有很好的理论和工程价值。