Physics-informed MPC slashes compute for drone control

Autonomous aerial vehicles demand control strategies that are both computationally efficient and robust in dynamic environments. A new paper from researchers Tayyab Manzoor, Yasir Ali, Yuanqing Xia, Lijie You, and Yan Wang introduces a physics-informed machine learning (PIML) approach to model predictive control (MPC) that balances these competing goals. The core innovation is a sparse, control-affine model identified via PIML, which embeds first-principles knowledge (e.g., Newtonian mechanics) while learning residual uncertainties from operational data. This yields a parsimonious yet interpretable representation of the vehicle's dynamics.

The identified model is integrated into a robust MPC scheme that uses high-order Runge-Kutta discretization for accurate predictions. Adaptive tube-based control tightens constraints based on the model's residual error, guaranteeing robust stability without excessive conservatism. Theoretical proofs establish recursive feasibility and stability. In simulations and real quadrotor experiments, the method drastically reduces computational load compared to standard nonlinear MPC and robust MPC with high-fidelity models, while achieving superior tracking performance over PID, nonlinear MPC, neural-network-based MPC, and fixed-tube robust MPC. This makes it ideal for resource-constrained aerial systems like delivery drones or surveying UAVs.