Sensor-Fusion Baselines and Robustness Envelopes for Real-Time UAV Obstacle Detection on Embedded Platforms

Abstract

While achieving effective obstacle identification and secure navigation under computational limitations, autonomous unmanned aerial vehicles (UAVs) rely on precise sensor fusion. This paper explores a wide-ranging benchmark and robustness analyses of eight artificial intelligence (AI) models over a multi-sensor UAV navigation dataset, including Light Detection and Ranging (LiDAR), GPS, IMU, and telemetry data. In this paper, standard machine-learning (logistic regression, SVC, k-NN, and Gaussian Naive Bayes) classifiers are compared with ensemble learners (Random Forest, XGBoost, and LightGBM) and a shallow MLP. The accuracy, F1-score, recall, AUC, calibration, inference latency, and memory footprint of each model are evaluated to measure their realistic baselines for onboard implementation. Boosting ensembles are well-suited to embedded UAV applications, reaching near-perfect obstacle-detection accuracy with high computational efficiency (XGBoost, LightGBM). Robustness envelopes constructed under LiDAR noise and sensor ablation scenarios reveal the operational boundaries of sensor-dominant regimes and highlight the importance of redundancy in sensor fusion. The results explained with SHAP show that the main decision variables are LiDAR distance and IMU dynamics. The framework provides repeatable baselines, measures of effectiveness, and safety-aware decision-making that can inform future efforts in energy-efficient and interpretable UAV autonomy.