Safe Reinforcement-Learning Navigation of Quadrotors in Procedurally Generated Forests: PX4-Calibrated Simulation, Layered Safety Filters, and Software-in-the-Loop Validation

Flying drones through dense, unmapped forests is extremely challenging because they must dodge tree trunks, branches, and moving hazards. To solve this, researchers developed a complete, reproducible navigation system tested in the PX4 Autopilots simulator. The system combines a fast, PX4-calibrated simulator, procedurally generated 3D forests and city environments, and a 7-channel LiDAR sensor. Instead of relying solely on an artificial intelligence policy—which tends to crash frequently on its own—the team implemented a layered safety shield and a Control Barrier Function (CBF) filter that adjusts the drone's path in real time, even accounting for moving obstacles. Comprehensive testing showed that these safety filters do the heavy lifting for collision avoidance, drastically reducing crash rates and allowing the drone to successfully complete complex autonomous missions without human intervention.