A Robust Machine Learning Method for Safe and Efficient UAS Integration

This study explores an automated approach to enhancing the robustness of machine learning (ML) models developed for drone collision avoidance, ensuring they meet critical safety and performance standards. The collision avoidance algorithms integrate ML with control theory, employing Control Barrier Function (CBF) and Graphical Neural Network (GNN). This research uses adversarial training to identify the worst-case scenarios within the uncertainty bounds for both sensor and process uncertainties. The efficacy of automatically identifying rare scenarios and strengthening model robustness through adversarial training is demonstrated in laboratory settings and validated through practical tests with micro aerial vehicles (MAVs).