Autonomous UAV swarm exploration in large-scale complex environment

Abstract

Autonomous unmanned aerial vehicles UAVs hold transformative potential for applications in complex environments, yet their widespread deployment is hindered by challenges in perception, planning, simulation, multi-agent coordination, and control. This thesis addresses these challenges through five key contributions, advancing the state of the art in UAV autonomy.

First, D-Map, a novel LiDAR-based occupancy mapping framework, is proposed to enhance computational and memory efficiency in large-scale environments. By leveraging depth image projection and a hybrid on-tree update strategy, the framework processes high-resolution LiDAR data efficiently while maintaining mapping accuracy. Its decremental property, enabled by LiDAR’s low false alarm rates, further optimizes resource usage. Benchmark and real-world experiments demonstrate its superiority in handling long-range, dense LiDAR measurements.

Second, a fully autonomous quadrotor system is developed to avoid dynamic, small obstacles (e.g., 9mm bars) in cluttered environments at speeds up to 5.5 m/s. The system integrates a 3D LiDAR and onboard processing for pose estimation, local mapping via time-accumulated point cloud KD-Trees, and kinodynamic A* trajectory planning—all running at 50 Hz. Indoor and outdoor trials validate its reliability in navigating complex scenarios, bridging the gap between high-speed flight and safety.

To address the scarcity of realistic simulation tools, MARSIM, a lightweight LiDAR simulator for UAVs, is introduced. By rendering depth images directly from real-world point cloud maps (e.g., forests, buildings), the simulator generates realistic LiDAR data, supports dynamic obstacles, and accommodates diverse LiDAR types (spinning/solid-state). Evaluations show it outperforms Gazebo in computational efficiency and accurately mirrors real-flight behavior. Ten high-resolution maps are provided, enabling diverse testing scenarios without dense mesh models.

For large-scale exploration, a multi-UAV system with a double-layer grid map (coarse/fine resolutions) is proposed. A partially scanned voxel method reduces inter-drone communication bandwidth while ensuring high-precision mapping. Real-world experiments, including nighttime mountain forests and underground parking lots, demonstrate five drones collaboratively scanning 10,000 m² in three minutes. The system’s efficacy is further highlighted in a search-and-rescue mission, with simulations confirming enhanced terrain adaptability and exploration efficiency.

Finally, the swashplateless-elevon actuation (SEA) for dual-rotor tail-sitter UAVs is proposed. By decoupling pitch (motor speed modulation) and yaw (elevons), the method mitigates control saturation and improves trajectory tracking, disturbance rejection, and takeoff stability compared to conventional approaches. Flight tests validate its robustness in hover, transition, and fixed-wing modes, even under high-speed airflow.

Collectively, this thesis advances UAV autonomy through innovations in perception, planning, simulation, swarm coordination, and control, validated across diverse real-world and simulated environments. The research results have greatly promoted the application ability of UAV and UAV swarm in the real complex environment, such as disaster response, environmental monitoring, and infrastructure inspection.