期刊
IEEE ROBOTICS AND AUTOMATION LETTERS
卷 7, 期 4, 页码 11735-11742出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/LRA.2022.3205442
关键词
Aerial systems: mechanics and control; motion control; reinforcement learning
类别
资金
- JSPS KAKENHI [JP19K04850]
- European Research Council (ERC) through European Union [679355]
- UKRI Trustworthy Autonomous Systems Node in Functionality [EP/V026518/1]
- European Research Council (ERC) [679355] Funding Source: European Research Council (ERC)
This letter experimentally investigated the reality gap effect in deep reinforcement learning for flight controllers of UAVs. The study focused on fixed-wing UAV pitch control in wind tunnel tests and compared three different training approaches. The results showed that the baseline approach was susceptible to the reality gap, while the high-fidelity modeling approach and the domain randomization approach successfully transferred to real tests. Additionally, the study found that the domain randomization controller maintained its performance even with configuration changes, highlighting its applicability to real environments with uncertainty.
Deep reinforcement learning has great potential to automatically generate flight controllers for uncrewed aerial vehicles (UAVs), however these controllers often fail to perform as expected in real world environments due to differences between the simulation environment and reality. This letter experimentally investigated how this reality gap effect could be mitigated, focusing on fixed-wing UAV pitch control in wind tunnel tests. Three different training approaches were conducted: a baseline approach that used simple linear dynamics, a high-fidelity modeling approach, and a domain randomization approach. It was found that the base line controller was susceptible to the reality gap, while the other two approaches successfully transferred to real tests. To further examine the controllers' capabilities to generalize, a variety of configuration changes were experimentally implemented on the UAV, such as increased inertia, extended elevator area, and aileron offset. While the high-fidelity controller failed to cope with these changes, the controller with domain randomization maintained its performance. These results highlight the importance of selecting appropriate sim-to-real transfer approaches and how domain randomization is applicable to fixed-wing UAV control with uncertainty in real environments.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据