Enhancing Neural Model Predictive Control for Agile Quadrotors

Model Predictive Control (MPC) has emerged as a powerful framework for real-time trajectory optimization in agile robotics, but its success hinges on the fidelity and speed of the underlying dynamics model. Recent advancements have shown that neural networks can model complex dynamic phenomena and be integrated into embedded MPC frameworks for quadrotor control. Building on the state-of-the-art Real-time Neural MPC (RTN-MPC) framework, we introduce two key improvements that significantly enhance performance and modeling flexibility.First, we develop a featurewise multi-layer perceptron (MLP) architecture tailored for multivariate time-series inputs, enabling the model to leverage full state and control histories for more expressive and efficient dynamics modeling. This innovation improves both prediction accuracy and computational efficiency. Second, we integrate a long short-term memory (LSTM) based model for capturing temporal dependencies in system behavior, demonstrating its strong generalization and control performance across aggressive flight trajectories.Through simulation experiments, we show that our models achieve superior tracking accuracy while maintaining real-time feasibility. These results highlight the potential of structured deep learning models for next-generation neural MPC pipelines in agile aerial robotics. This talk will discuss the architecture design, integration into MPC, and insights from simulation studies, setting the stage for future real-world deployment.