Autonomous navigation for MAVs : achieving agile, high-speed, and safe flight

Abstract

Micro air vehicles (MAVs) are capable of high-speed, agile maneuvers, making them ideal for time-critical applications like search and rescue or disaster relief. However, achieving agile, high-speed, and safe autonomous navigation for MAVs is challenging, requiring reduced vehicle size and weight for maneuverability, strong sensing capabilities to detect obstacles at a distance, and advanced planning and control algorithms that balance speed and safety. This thesis develops efficient mapping and planning algorithms for LiDAR-equipped MAVs, leveraging the advanced sensing capabilities of LiDAR and the agile performance of MAVs. The key contributions are as follows:

First, this thesis introduces a robot-centric occupancy mapping framework called ROG-Map. ROG-Map uses a uniform grid structure to maximize map updating and querying efficiency, alongside a zero-copy map sliding strategy to minimize memory overhead, enabling large-scale missions. A novel incremental obstacle inflation method reduces computational costs to just 1% of traditional batch inflation methods. Evaluations using benchmark datasets and real-world experiments show ROG-Map's superior computational efficiency and overall performance.

Second, this thesis addresses the challenge of online whole-body motion planning (SE(3) planning) for MAVs to perform agile movements, such as flying through narrow gaps in unknown, cluttered environments. A novel multi-resolution search algorithm actively identifies narrow regions requiring SE(3) planning. This is combined with a planning problem decomposition strategy that applies SE(3) planning only in these regions, while using computationally efficient R3 planning elsewhere. Simulations show this method reduces computation time by one to several orders of magnitude compared to state-of-the-art (SOTA) baselines, without sacrificing success rates. Real-world experiments demonstrate the first successful implementation of online whole-body motion planning for a quadrotor MAV in unknown environments using only onboard sensing and computation.

Finally, this thesis presents a comprehensive solution for high-speed and safe MAV navigation. We first introduce BubblePlanner, which combines a novel sample-based corridor generation technique with an efficient differentiable trajectory optimization framework. Evaluations in both simulations and real-world experiments demonstrate its superior computational efficiency and high-quality trajectory generation. To further enable safety-assured high-speed flights, the thesis presents the Safety-assUred high-Speed aErial Robot (SUPER) system. SUPER generates two trajectories during each re-planning cycle: one in known-free spaces to ensure safety and another in both known-free and unknown spaces to maximize speed. Notably, both trajectories are generated directly from LiDAR point clouds, enabled by a novel on-point-cloud known-free area extraction theorem. Additionally, a new switching time optimization method is proposed to determine the optimal transition between the two trajectories, maximizing flight speed while ensuring safety. Compared to baseline methods, SUPER reduces failure rates by 35.9 times, achieves higher flight speeds, and requires only half the planning time. Real-world experiments validated the system’s performance, demonstrating safe autonomous flights at speeds exceeding 20 m/s, successfully avoiding thin branches and navigating narrow passages inaccessible to state-of-the-art commercial products. SUPER’s versatility was further demonstrated across tasks such as target tracking, autonomous exploration, and waypoint navigation, marking a significant milestone in transitioning high-speed, robust MAV systems from laboratory research to real-world deployment.