Meta-Learning for Quadcopter Control (Fall 2026)

This repository provides the code and experimental framework for learning neural network–based quadcopter controllers via meta-learning. The goal is to enable rapid adaptation to previously unseen wind disturbances in hover and regulation tasks using only a small number of samples. The project applies Model-Agnostic Meta-Learning (MAML.pdf) to learn a neural feedback policy initialization that can be efficiently fine-tuned to new disturbance conditions via a small number of gradient updates.

Project report: DECODE_Report_Goffin.pdf

---

Key Results

• The neural network controller matches a classical LQR-like behavior on basic regulation tasks.
• Standard neural network training can struggle when data are unbalanced across wind conditions.
• MAML yields faster and more reliable few-shot adaptation to unseen wind conditions, on both linearized and nonlinear dynamics.

---

Future Work

• Extend the framework to higher-fidelity nonlinear models to further reduce the sim-to-real gap.
• Incorporate formal stability guarantees into the learning-based controller.
• Validate the approach on real-world quadcopter hardware.

---

Installation

git clone https://github.com/DecodEPFL/MetaLearning-Control-Quadcopters.git
cd MetaLearning-Control-Quadcopters

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Acknowledgements

Semester Project, EPFL DECODE Lab, Fall 2026
Author: Cyril Goffin
Supervision: Daniele Martinelli & Nicolas Kirsch
Professor: G. Ferrari Trecate

---