Data-Efficient Deep Reinforcement Learning for Attitude Control of Fixed-Wing UAVs: Field Experiments.

Area of Science:

• Aerospace Engineering
• Control Systems
• Machine Learning

Background:

• Fixed-wing unmanned aerial vehicle (UAV) attitude control presents challenges due to nonlinear dynamics, actuator limits, and motion coupling.
• Current linear control autopilots have performance limitations with complex UAV dynamics.

Purpose of the Study:

• To investigate the efficacy of deep reinforcement learning (DRL) for UAV attitude control.
• To demonstrate DRL's ability to learn optimal control laws directly from nonlinear dynamics.

Main Methods:

• Deep reinforcement learning (DRL) was employed to discover control laws through system interaction.
• Controllers were trained in simulation and then deployed on a real UAV for flight testing.
• Learned controllers were quantitatively and qualitatively compared against the ArduPlane PID autopilot.

Main Results:

• DRL successfully learned UAV attitude control using only 3 minutes of flight data.
• The DRL controller demonstrated performance comparable to the state-of-the-art ArduPlane PID controller.
• Actuation delay and diverse simulated dynamics were critical for successful real-world transfer.

Conclusions:

• DRL offers a powerful method for UAV attitude control, adept at handling complex nonlinear dynamics.
• The learned DRL controller achieved robust performance on a real UAV without online adaptation.
• This research highlights DRL's potential to advance UAV autopilot capabilities.