Trajectory Tracking Controller for Quadrotor by Continual Reinforcement Learning in Wind-Disturbed Environment.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Quadrotor trajectory tracking accuracy degrades in dynamic wind fields, challenging conventional controllers.
• Existing data-driven methods face catastrophic forgetting, limiting environmental adaptability.
• Robust control in variable wind conditions is crucial for complex environmental missions.

Purpose of the Study:

• To develop a reinforcement learning framework with continual adaptation for robust quadrotor tracking in dynamic wind fields.
• To address the limitations of conventional and data-driven approaches in handling wind disturbances.
• To enhance the environmental adaptability and tracking performance of quadrotors.

Main Methods:

• A continual reinforcement learning framework integrating continual backpropagation and reinforcement learning.
• Initial training in wind-free conditions followed by dynamic neuron resetting via utility assessment.
• A multi-objective reward function for improved training precision and efficiency.
• Validation using the Gazebo/PX4 simulation platform with stepwise and stochastic wind variations.

Main Results:

• Demonstrated reduction in root mean square error of trajectory tracking compared to the standard Proximal Policy Optimization (PPO) algorithm.
• Successfully resolved the plasticity loss problem in deep reinforcement learning through structured neuron resetting.
• Significantly enhanced continual adaptation capabilities of quadrotors in dynamic wind fields.

Conclusions:

• The proposed framework offers a robust solution for quadrotor trajectory tracking in challenging, time-varying wind disturbances.
• Continual adaptation through structured neuron resetting maintains network plasticity and improves performance.
• This approach advances the reliability of quadrotors in complex environmental missions.