Neural-Augmented Planning and Tracking Pilots for Maneuvering Multibody Dynamics

We propose a methodology for extending the applicability of multibody-based comprehensive analysis codes to the maneuvering regime, with specific application to the flight of rotorcraft vehicles. Maneuvers are here mathematically described in a concise yet completely general form as optimal control problems, each maneuver being defined by a specific form of the cost function and by suitable constraints on the vehicle states and controls. In principle, by solving the maneuver optimal control problem, one could determine the trajectory and the control time histories that steer the vehicle model, while minimizing the cost and satisfying the constraints. Unfortunately, optimal control problems are prohibitively expensive to solve for detailed comprehensive models of rotorcraft vehicles denoted by a large number of structural degrees of freedom and possibly sophisticated aerodynamics. In order to make the problem computationally tractable, our formulation makes use of two models of the same vehicle. A coarse level flight mechanics model is used for solving the trajectory optimal control problem. Being based on a reduced model of the vehicle with only a few degrees of freedom, the resulting non-linear multi-point boundary value problem is computationally feasible. Next, the fine scale comprehensive model is steered in closed loop, tracking the trajectory computed at the flight mechanics level using a receding horizon model predictive controller. This amounts to a standard time marching problem for the comprehensive model, which is therefore also computationally feasible. The flight mechanics model is iteratively updated for ensuring close matching of the trajectories flown by the two models, by resorting to a neural adaptive element. This two-level procedure enables the simulation using comprehensive models of arbitrary complexity of maneuvers of possibly long duration, with general constraints on the vehicle inputs and outputs. The new procedures are demonstrated with the help of numerical applications