Robust Reinforcement Learning Control Strategy for Vision-Based Ship Landing of Vertical Flight Aircraft
Abstract
This study discusses a fully autonomous vertical flight aircraft ship landing procedure in presence of wind disturbances. The proposed study closely follows the established Navy helicopter ship landing procedure wherein the pilot utilizes the ship as the visual reference for long-range tracking; however, upon coming closer, the pilot follows a gyro-stabilized horizon bar installed on most Navy ships to approach and land vertically independent of deck motions. This was accomplished by developing a unique vision system and a hybrid control system validating its performance in simulations and flight tests.
The vision system serves the purpose of a pilot’s eye by obtaining the visual information required for a safe approach and landing. The vision system can be engaged from 250 meters away from the landing pad, initially utilizing machine learning strategies to detect the ship for long-range tracking and switches to a unique combination of classical computer vision techniques to detect the horizon bar to precisely estimate the aircraft position and orientation relative to the bar during the final approach and landing. The distance and attitude estimations were validated using the measurements from an accurate 3D motion capture system (VICON), which demonstrated sub-centimeter and sub-degree accuracy.
Finally, a hybrid control system is developed to control the aircraft using the perceived visual
information. The hybrid control system is a combination of a non-linear controller and a Deep
Reinforcement Learning (RL) controller. The non-linear controller demonstrated robust tracking
capability even in presence of estimation noise and varying time delays between successive control actions. The RL controller is developed exclusively for disturbance rejection. When conducted flight testing in presence of 5 m/s wind, the RL controller shows a 100% reduction in drift and a 10 times faster rate of correction compared to a conventional control system. The vision and hybrid control system were implemented on a quadrotor UAV and extensive flight tests were conducted to demonstrate accurate tracking in challenging conditions and safe vertical landing on a sub-scale ship platform undergoing 6 degrees of freedom deck motions.
Citation
Saj, Vishnu (2021). Robust Reinforcement Learning Control Strategy for Vision-Based Ship Landing of Vertical Flight Aircraft. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /196365.