Deep Reinforcement Learning Control for Quadrotor

This paper investigates a physics-informed deep reinforcement learning approach for the agile tracking and obstacle avoidance control of quadrotors. Traditional control methods often struggle to account for the strong aerodynamic coupling and nonlinear dynamics inherent in high-speed maneuvers. While deep reinforcement learning offers a promising alternative, standard black-box approaches frequently suffer from low sample efficiency, limited physical interpretability, and slow strategy convergence. To overcome these limitations, a physics-informed deep reinforcement learning framework built upon the proximal policy optimization algorithm is proposed. By incorporating Newton-Euler dynamics, a physics-informed neural network architecture is integrated where the drone's governing equations are embedded as physical residual terms within the total loss function. A robust state space that accounts for measurement noise and a normalized action space are arranged to respect motor physical constraints. Furthermore, a dense reward function is formulated to balance tracking precision, flight stability, and task safety. To ensure continuous and smooth control commands, Kullback-Leibler divergence regularization and momentum-based smoothing techniques are implemented. Simulation results show that the proposed algorithm can effectively improve the tracking accuracy and obstacle avoidance safety of quadrotors.