Aerodynamics Co-Design

Gradient-based nested co-design of aerodynamic shape and flight trajectory for agile glider manoeuvres such as perching and soft landing.

This repository implements a differentiable bilevel optimization pipeline in which the outer loop optimizes the airfoil or wing shape, while the inner loop solves a constrained optimal control problem (OCP) for the vehicle dynamics. The dynamics use a physically informed neural network surrogate to predict aerodynamic coefficients.

The repository currently contains two related aerodynamic formulations:

• a 2D section-based formulation, which is the methodology documented and validated in our paper, and
• a 3D lifting-line-coupled extension, implemented in the codebase as a subsequent development.

Repository structure

The repository is organized around three complementary layers:

glider_optimization/
├── glider_optimization/   # core package
│   ├── blocks/            # differentiable pipeline blocks
│   ├── utils/             # lifting-line, geometry, and utility functions
│   ├── config.py
│   ├── runner.py
│   └── main.py
├── conf/                  # run configurations
└── methods/               # aerodynamics background and modeling notes

The structure of the pipeline is designed to be modular. It is possible to change the control problem by editing glider-optimization/blocks/ocp.py, or even skip it entirely, resulting in a fixed-point shape optimization. In that case, the evaluation function defined in glider-optimization/blocks/evaluation.py should be updated accordingly.

Installation

Clone the repository and install it in editable mode:

pip install -e .

Quick start

Run a configuration:

glider-opt --config conf/test.yaml --run-name demo_run

Key configuration knobs

A few settings determine most of the behavior:

• neuralFoilSampling.use_3d_llt

  • false: use the 2D sectional formulation
  • true: use the 3D lifting-line-coupled formulation

• neuralFoilSampling.n_samples

  • number of aerodynamic samples used to build the local reduced-order model

• neuralFoilSampling.neuralFoil_size

  • NeuralFoil model size used for sectional predictions

• reducedModel.chebyshev_degree

  • degree of the Chebyshev surrogate used inside the OCP

• evaluation.mode

  • task objective, for example Perching or SoftLanding

• plane.wing.*

  • finite-wing geometry definition, including span stations, chord, sweep, twist, and dihedral

Citation

If you use this repository in academic work, please cite the accompanying paper:

@article{affinita2026gradient,
  title={Gradient-based Nested Co-Design of Aerodynamic Shape and Control for Winged Robots},
  author={Affinita, Daniele and Xu, Mingda and Gherardi, Beno{\^\i}t Valentin and Fua, Pascal},
  journal={arXiv preprint arXiv:2603.06760},
  year={2026}
}

License

This repository is licensed under Apache 2