Towards autonomous aerial swarms : from sensor calibration to state estimation

Abstract

Aerial swarm systems exhibit significant potential across a range of applications, including cooperative exploration, target tracking, and search and rescue, primarily due to their exceptional capabilities for team coordination. However, the collaborative execution of these complex tasks poses substantial challenges, necessitating advanced sensing capabilities for environmental perception, robust and precise sensor calibration methodologies, and efficient algorithms for ego and mutual state estimation that effectively balance performance and accuracy. This thesis presents the development of robust, user-friendly sensor calibration tools and fully decentralized swarm state estimation algorithms, thereby enhancing the ability of LiDAR-equipped aerial swarm systems to successfully undertake diverse missions in intricate scenarios. The key contributions are as follows:

First, this thesis introduces LI-Init: a comprehensive, robust, and real-time LiDAR-inertial initialization algorithm that calibrates the temporal offset and extrinsic parameter between LiDARs and inertial measurement unit (IMUs), and also the gravity vector and IMU bias. This is achieved by aligning the states estimated from LiDAR measurements with that measured by IMU. LI-Init autonomously calibrates spatial-temporal offsets, providing reliable initial states without the need for target markers or additional sensors, enabling a LiDAR-inertial odometry to run on a customized sensor setup without any dedicated prior calibration or hardware setup. Experimental evaluations acrossvarious LiDAR systems and LiDAR-inertial configurations demonstrate the robustness, adaptability, and efficiency of the proposed initialization method.

Second, this thesis further enhances the efficiency of radar-inertial odometers and addresses, to some extent, the issue of LIO being unable to run in real-time on low-power embedded computing devices. A novel LiDAR-inertial odometry (LIO) framework is designed to achieve robust and efficient localization while significantly reducing computational cost. A degeneration-aware batch selector is introduced to dynamically adjust the number of points for scan registration based on real-time assessment of environmental constraints. This algorithm ensures robust localization in degenerate scenarios and avoids unnecessary computation in well-constrained environments. Extensive experiments on public datasets, UAV platforms, and resource-constrained embedded devices demonstrate that the proposed LIO (termed as ELIO) achieves comparable or better localization accuracy than state-of-the-art methods.

Third, this thesis preliminarily addresses the challenge of swarm state estimation for LiDAR-equipped aerial swarm systems by introducing Swarm-LIO, a decentralized state estimation method designed specifically for aerial swarm operations. In this framework, each drone autonomously performs precise ego-state estimation and exchanges ego-state and mutual observation information via wireless communication. Additionally, a novel three-dimensional (3D) LiDAR-based method for drone detection, identification, and tracking is proposed to enhance the acquisition of observations concerning teammate drones. The mutual observation measurements are then tightly coupled with IMU and LiDAR data, facilitating the joint real-time and accurate estimation of both ego-states and mutual states.

Finally, this thesis introduces a comprehensive solution, designated as Swarm-LIO2, for LiDAR-inertial state estimation in aerial swarm systems. Swarm-LIO2 builds upon the foundational framework established by Swarm-LIO, which delineated the general principles of swarm LiDAR-inertial odometry. In contrast to its predecessor, Swarm-LIO2 incorporates several critical enhancements. To optimize the swarm initialization process, we employ factor graph optimization techniques for expedited teammate identification and global extrinsic calibration. This advancement significantly diminishes the required number of flights for a swarm of N unmanned aerial vehicles (UAVs) from O(N) to O(1). Furthermore, we introduce an innovative state marginalization strategy alongside a LiDAR degradation evaluation method, both designed to alleviate computational burdens and enhance swarm scalability. Extensive simulations and real-world experiments, including inter-UAV collision avoidance, target tracking , and cooperative payload transportation, demonstrate the superior versatility and practical significance of Swarm-LIO2, marking a substantial milstone in the field of state estimation for LiDAR-equipped aerial swarm systems