Train of Autonomous Aerial Vehicles for Subterranean Exploration With SLAM Capabilities
Nicolas Blanchard, Amber Parker, Abhijit Mahalanobis, Sergey V. Shkarayev · 2025
Train of Autonomous Aerial Vehicles for Subterranean Exploration with SLAM Capabilities Nicolas Blanchard , Amber Parker , Abhijit Mahalanobis , Sergey Shkarayev The University of Arizona, Tucson, Arizona, 85719, USA This work realizes a system of drones for the exploration of caves and lava tubes. The presented system consists of lightweight quadcopters that employ Visual-Inertial Odometry for real-time localization and Time of Flight occupancy grid mapping for real-time 3D mapping, yielding similar results to the online SLAM solution but without the need for a 360 degree 3D LiDAR scanner. A waypoint-based Leader-Follower algorithm and a custom MAVLink-based ground station are deployed to control the system and direct the swarm of UAVs in a train configuration. Software-in-the-Loop simulations and complete system flight tests in an artificial cave environment area conducted to evaluate system communication, navigation, and 3D mapping ability. Robust leader-follower configuration, 3D occupancy mapping at resolutions of 0.05 meter, and accurate localization in GPS-denied and low-light conditions are all demonstrated. Nomenclature SLAM = Simultaneous Localization and Mapping VIO = Visual-Inertial Odometry ToF = Time of Flight UAV = Unmanned Aerial Vehicle PX4 = Open source flight control software GPS = Global positioning system x_VIO = x-axis position estimate of Visual-Inertial Odometry y_VIO = y-axis position estimate of Visual-Inertial Odometry z_VIO = z-axis position estimate of Visual-Inertial Odometry 〖∠〖roll〗_VIO〗_ = Roll orientation estimate of Visual-Inertial Odometry 〖∠〖pitch〗_VIO〗_ = Pitch orientation estimate of Visual-Inertial Odometry 〖∠〖yaw〗_VIO〗_ = Yaw orientation estimate of Visual-Inertial Odometry Introduction Autonomously operating unmanned aerial vehicles (UAVs) are key for exploring unknown spaces in a variety of applications. Search and rescue missions, building exploration, surveying, reconnaissance, and even combat operations all benefit from a robust system of vehicles capable of remote exploration. Two additional comparable environments are caves and lava tubes. Caves and lava tubes are important because they provide a unique environment for preserving microbial life. Orbiters have discovered dark pits on the surface of Mars that resemble caves or lava tubes. By analogy with Earth, caves on Mars may contain signatures of past microbial life and become valuable resources of lava-flow thermodynamics and hydrodynamics gravity. Figure 1 shows a candidate for a cave entrance on mars captured by the HiRISE imager onboard NASA’s Mars Reconnaissance Orbiter. The image was captured on Aug. 16, 2020. Fig. 1 Possible entrance to lava tube near Arsia Mons on Mars. Caves and lava tubes, along with any indoor/dense environment mentioned above, are unknown environments and present several key challenges. The primary considerations are lack of GPS or other global positioning information, low light conditions, intricate branching structure, loss of RF communication signals around corners, and rock formations and other obstacles. In addition, due to the remote nature of extraterrestrial caves, any system that aims to successfully explore and characterize the subterranean surroundings must respond to stimuli and navigate itself with minimal operator input (autonomous operation). Simultaneous Localization and Mapping (SLAM) is the state-of-the-art approach for UAV localization in an unknown environment. Unlike approaches relying on pure odometry, SLAM is preferred for its ability to correct for trajectory drifts through loop closure [1]. In addition, certain SLAM approaches build 3D maps of the surrounding environment, yielding knowledge of the features and obstacles that a vehicle has encountered. SLAM techniques have been refined for decades, with many different flavors being introduced. This work focused on the applications of visual SLAM, a form of SLAM that uses visual cameras as the primary data