Pareto-Adaptive Robust Neural Control With Decision-Making for Multicopter

A flight control system must balance tracking accuracy, control smoothness, and energy use under nonlinear dynamics, disturbances, actuator dynamics, and delays. Existing schemes (e.g., LQR, MPC) typically treat the problem as single-objective with fixed weights, provide limited guarantees for multi-objective trade-offs, and often neglect actuator dynamics, delays, or automated selection of weights—making consistent performance under uncertainty difficult. This paper presents an adaptive Neural-network Pareto-based Optimal Control (NPOC) framework that formalizes a weighted four-objective cost—energy usage, state deviation, control smoothness, and time-weighted tracking error. We prove the existence of optimal solutions within the weighted multi-objective formulation and derive a new state-feedback control law with closed-loop stability established via Lyapunov analysis. Actuator dynamics and signal delays are explicitly modeled within the control synthesis. A multilayer-perceptron decision maker, trained on simulation-augmented data, maps desired performance metrics (RMSE and energy) to weight selections, enabling automatic adaptation of the controller. Validation on a nonlinear quadcopter with wind disturbances, sensor noise, and 50% mass variation shows improved tracking accuracy, smoother inputs, and favorable energy usage compared with LQR and MPC. Monte Carlo trials over multiple initial conditions and ANOVA tests (p-values above significance thresholds) confirm robustness and statistical reliability. These results demonstrate a mathematically grounded and adaptive control methodology for multicopters operating under complex nonlinear dynamics and uncertainty.