Real-time embedded safe landing area detection using flash LiDAR and deep neural networks for aerial vehicles

We present a landing-site evaluation system that uses Flash LiDAR point clouds and deep neural networks (DNNs) to identify safe landing zones and hazards for a wide range of aerial vehicles. The DNN approach is trained entirely on high-fidelity synthetic data generated in simulation and verified with the NASA Morpheus dataset collected with an Advanced Scientific Concepts (ASC) Global Shutter Flash LiDAR (GSFL).

Our method integrates 3D terrain stitching for wide-area digital elevation map (DEM) reconstruction with learned hazard assessment models to predict safe landing areas in the surveyed terrain. All components, including the DNN inference and geometric processing, run at real-time sensor frame rates on an embedded processor, satisfying onboard autonomy requirements comparable to the NASA Morpheus system.

We further compare our DNN-based predictions with a classical Safe Landing Area Determination (SLAD) algorithm to assess performance, trade-offs, and robustness. Finally, we demonstrate that our simulation pipeline effectively bridges the sim-to-real gap, enabling reliable transfer of trained models to field-deployable systems. This work lays the groundwork for advanced onboard autonomous landing decision systems for manned and unmanned aerial vehicles across terrestrial and planetary missions.