Learning Quadrotor Dynamics Using Feed-Forward Neural Networks

This paper presents a neural network-based approach for modeling quadrotor dynamics. Two feed-forward neural networks (FFNNs) are designed to predict linear and angular accelerations using a high-resolution flight dataset encompassing diverse flight regimes and maneuvers. The data-driven method addresses limitations of traditional physics-based models by effectively capturing non-linearities and real-world complexities. Results show the proposed FFNNs achieve accurate predictions, highlighting their potential to enhance quadrotor modeling and facilitate advanced control strategies.