General Purpose Low-Level Reinforcement Learning Control for Multi-Axis Rotor Aerial Vehicles.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Existing reinforcement learning controllers for multirotor drones are often model-specific, requiring extensive retraining for different vehicles.
• This limitation hinders the adaptability and practical application of reinforcement learning in drone control.

Purpose of the Study:

• To develop a multipurpose, model-free reinforcement learning control structure for low-level multirotor unmanned aerial vehicles (UAVs) using neural networks.
• To create a control policy that can be applied to various multirotor configurations without model-specific or parameter-specific constraints.

Main Methods:

• A 6-degree-of-freedom dynamic model is employed, integrating acceleration-based control from a policy neural network.
• An end-to-end neural network learns maneuvers from fused sensor states to acceleration commands, utilizing on-board sensors and motion capture data.
• A multisensory fusion framework compensates for time delays and low update frequencies from the motion capture system.

Main Results:

• The trained control policy, implemented with an improved algorithm, is directly mapped to actuators and applied to various multirotors without expert demonstration.
• The algorithm successfully demonstrated control capabilities in hovering and tracking tasks for both quadrotor and hexrotor configurations.
• The same policy effectively stabilized both quadrotor and hexrotor drones under random initial states in simulations and experiments.

Conclusions:

• The proposed reinforcement learning control structure offers a versatile and efficient solution for low-level multirotor UAV control.
• The model-free approach significantly reduces training time and enhances adaptability across different multirotor platforms.
• This research validates the effectiveness of the developed policy for stable flight control in diverse multirotor configurations.