What is it about?
Autopilot systems allow aircrafts to be controlled without the help of a pilot. Autopilot systems are an important area of research for drone technologies. Traditional autopilot systems consist of an "inner loop" and an "outer loop." The inner loop provides stability and control. On the other hand, the outer loop is for mission-level control, such as controlling navigation. Traditional controllers of autopilot systems work well under stable conditions. However, their performance is inadequate under harsh and variable conditions. "Reinforcement learning (RL)," a type of machine learning, can be used to improve the performance of autopilot controllers. Thus far, RL has focused mainly on mission-level control, not inner loop control. This study tested the performance and accuracy of an inner loop control that has been trained with RL algorithms. First, the authors developed a simulation environment, that feels close to a real flight scenario. They then used it to train the flight controller of a drone. They then compared the performance of the controller to a traditional controller. They found that along with accurate tracking, the controller trained with RL also reduced error by 44%.
Featured Image
Photo by Iewek Gnos on Unsplash
Why is it important?
Autopilot systems must be able to adapt to changes in the flight environment. Maintaining the vehicle's direction or "attitude control" under extreme conditions is a major challenge. The findings presented in this paper can be a first step towards exploring the use of RL for smart flight control. It also shows ways to improve the accuracy of attitude control. KEY TAKEAWAY: This study shows how RL training can be used to improve smart flight control. It will enable scientists and engineers to improve the quality of attitude control in autopilot systems.
Read the Original
This page is a summary of: Reinforcement Learning for UAV Attitude Control, ACM Transactions on Cyber-Physical Systems, April 2019, ACM (Association for Computing Machinery),
DOI: 10.1145/3301273.
You can read the full text:
Resources
Neuroflight...No longer theoretical, we are there
YouTube video
How to Train Your Quadrotor: A Framework for Consistently Smooth and Responsive Flight Control via Reinforcement Learning
This work addresses key shortcomings in the use of RL for smooth control of quadrotors. It closes the gap between good results in simulated flight and reliable transfer to real aircraft.
Regularizing Action Policies for Smooth Control with Reinforcement Learning
This work rethinks the role of reward signals vs. policy regularization to define a divide-et-impera approach to trade off control aggressiveness and smoothness. The new approach significantly widens the applicability of RL to tackle control problems to a broader class of robotic systems.
Contributors
Be the first to contribute to this page