AirPilot Drone Controller: Enabling Interpretable On-the-Fly PID Auto-Tuning via DRL

Precision, speed, and stability are essential for safe Unmanned Aerial Vehicle (UAV) flights in dynamic environments, where mission objectives can vary from minimizing energy consumption to maximizing speed and precision. Traditional Proportional Integral Derivative (PID) controllers, while popular, struggle with the nonlinear dynamics of wind and complex drone systems. Manual tuning of PID gains is time-consuming and expertise-dependent. This paper introduces the AirPilot, a nonlinear PID controller enhanced with Deep Reinforcement Learning (DRL) using Proximal Policy Optimization (PPO). The AirPilot combines the effectiveness of PID control with the adaptability and optimization capabilities of DRL, making it well-suited for modern, variable mission requirements. We trained our DRL agent using a COEX Clover autonomous drone within a simulator and implemented it in a real-world lab setting, significantly improving upon traditional control metrics. The AirPilot reduced navigation error of the default PX4 PID position controller by $\mathbf{9 0 \%}$, increased navigation speed by $21 \%$, and decreased both settling time and overshoot by $17 \%$ and $\mathbf{1 6 \%}$, respectively. Additionally, integrating a ${\$}$20,000 Vicon tracking system enhanced positioning accuracy to within 1 mm. We further developed and implemented a three-dimensional A* path planner for optimal collision-free navigation. This paper marks a pioneering step in applying DRL-based controllers to actual drones.