Collaborative Infrastructure-Free Aerial–Ground Robotic System for Warehouse Inventory Data Capture

Abstract

Highlights: What are the main findings? A fully autonomous, infrastructure-free UAV–UGV system for warehouse inventory isdesigned and validated: the UGV performs global localization via EKF + scan-to-mapusing only a 2D floor plan, while the UAV uses ArUco-based relative localization toautonomously take off, follow the UGV, and execute precision landings; results arevalidated in simulation and real-world trials. The proposed UGV localization yields consistently low translational error and the lowestorientation RMSE across scenarios, outperforming AMCL and remaining stable duringa 45-min warehouse run where AMCL diverged; the UAV reliably follows and landsusing a lightweight controller. What are the implications of the main findings? Enables rapid, low-cost deployment in GPS-denied, symmetric warehouses withoutprior 3D mapping or external infrastructure, remaining robust to shelf-occupancychanges and dynamic clutter. Provides a practical path toward scalable cycle counting by coordinating UGV motionwith UAV shelf scanning; the open-source implementation facilitates reproduction andadoption by industry and academia. Efficient and reliable inventory management remains a challenge in modern warehouses, where manual counting is time-consuming, error-prone, and costly. We present an autonomous aerial–ground system for warehouse inventory data capture that operates without external infrastructure or prior mapping operations. A differential-drive unmanned ground vehicle (UGV) performs global localization and navigation from a simple 2D floor plan via 2D LiDAR scan-to-map matching fused in an Extended Kalman Filter. An unmanned aerial vehicle (UAV) uses fiducial-based relative localization to execute short, autonomous take-off, follow, precision landing, and close-range imaging of high shelves. By ferrying the UAV between aisles, the UGV extends the UAV’s effective endurance and coverage, limiting flight to brief, high-value segments. We validate the system in simulation and real environments. In simulation, the proposed localization method achieves higher accuracy and consistency than AMCL, GMapping, and KartoSLAM across varied layouts. In experiments, the UAV reliably follows and lands on the UGV, producing geo-referenced imagery of high shelves suitable for downstream inventory recognition.