Autonomous Navigation for Object Capture with Multicopters

拟议项目的目标是开发郊区多飞行器自主导航的新方法。从第一个资助期的结果开始，将开发先进的直升机能力。直升机(在P2中)上的同心测绘使其也有可能规划机上的同心导航。P2中的对象检测允许相对于这些对象进行导航。对动态障碍的感知(在P2中)将是预见性规划的基础。启动和着陆将自动进行。在不同的时间尺度上追求导航目标。任务规划中的慢任务就是根据用户的需求来创建飞行任务。结果是一个姿势序列，主传感器，一个高分辨率的相机，应该在这个位置捕捉图像。在执行任务时，从直升机的当前姿态(从P1)到下一视图姿态，以中频率规划3D导航路径。规划者将同时优化多个标准：避开障碍物、保持沟通和本地化、风力大小和控制成本。基于P2中创建的以自我为中心的障碍物地图，局部规划者将生成具有高速3D避障路径的本地规划者。通过这种方式，直升机将能够对外部干扰做出快速反应，特别是在其附近探测到风和障碍物。为了在星载PC机有限的计算资源下进行高速率的规划，多分辨率方法将被提出。对于移动障碍物的规划，除空间离散化外，还应在时间维度上采用多分辨率。对于考虑直升机飞行动力学的相对导航，应采用多分辨技术加速快速模型预测控制。我们还致力于针对子系统丢失的健壮性，即传感器、通信或马达的丢失。对于所有这些情况，我们将制定适当的行为来克服损失(例如，战斗到基站或GNSS卫星的直接视线区域)，或者至少允许安全着陆。最后，我们的目标是从人类专家那里学习导航策略，并将它们转移到新的情况下。为此，我们将使用逆强化学习方法来学习代价函数。

Objective of the proposed project is the development of new methods for the autonomous navigation of multicopters in sub-urban areas. Starting from the results of the first funding period, advanced copter capabilities will be developed. Allocentric mapping onboard the copter (in P2) makes it possible to also plan allocentric navigation onboard. The detection of objects in P2 allows for navigation relative to these. The perception of dynamic obstacles (in P2) will be the basis for anticipatory planning. Start and landing will be made automatic. Navigation goals are pursued on different time scales. Task of the slow mission planning is the creation of a flight mission based on the requirements of the user. The result is a sequence of posed at which the main sensor, a high-resolution camera, shall capture images. When executing the mission, 3D navigation paths are planned with medium frequency from the current pose of the copter (from P1) to the next view pose. The planner will optimize multiple criteria simultaneously: the avoidance of obstacles, the maintenance of communication and localization, wind strength, and control costs. Based on the egocentric obstacle map created in P2, a local planner will generate with high rate 3D obstacle-avoiding paths. In this way, the copter will be able to react quickly on external disturbances, in particular wind and obstacle detections in its vicinity. In order to plan with high rates with the limited computational resources of the onboard PC, multi-resolution methods will be advanced. In addition to the spatial discretization, multiresolution shall be used in the time dimension for planning with moving obstacles. For the relative navigation under consideration of the copter flight dynamics, fast model predictive control shall be accelerated by multiresolution techniques. We also aim at robustness against the loss of sub-systems, i.e. of sensors, communication, or motors. For all of these cases, we will develop suitable behaviors to overcome the loss (e.g. fight into an area of direct sight to the base station or GNSS satellites) or at least allow for a safe landing. Finally, we aim at learning navigation strategies from human experts and their transfer to novel situations. To this end, we will learn cost functions with inverse reinforcement learning methods.