Gradient-based Nested Co-Design of Aerodynamic Shape and Control for Winged Robots

A research team from EPFL has published a novel AI framework that tackles a fundamental challenge in aerial robotics: designing robots whose physical form and flight control software are perfectly matched for specific, complex tasks. Traditional sequential design—first the wing, then the flight plan—is inherently suboptimal because aerodynamics and motion are nonlinearly intertwined. The team's solution is a general-purpose, gradient-based, nested co-design framework that optimizes both the aerodynamic shape and the motion planner simultaneously.

At its core, the method uses a neural network as a surrogate model to predict complex aerodynamic forces under various flight conditions, which are then fed into an optimal control problem solver. This approach overcomes the simplifying assumptions that have limited prior co-design methods. The researchers validated their framework on two demanding dynamic tasks for fixed-wing gliders: executing a precise perching maneuver and achieving a short landing. The AI-optimized designs significantly outperformed those from a traditional evolutionary algorithm baseline, and crucially, achieved these results in a small fraction of the computation time, demonstrating both superior performance and practical efficiency.