Bayesian Optimization for Fast Radio Mapping and Localization with an Autonomous Aerial Drone

TL;DR

This work addresses autonomous drone localization using narrowband radio signals by combining Gaussian Process-based radio mapping with Bayesian optimization to select informative sampling locations. The radio field $f(\cdot)$ is modeled with a GP prior $f(\cdot) \sim \mathcal{GP}(m(\cdot), k(\cdot,\cdot))$, yielding a posterior mean $\mu(\mathbf{x})$ and uncertainty $\sigma(\mathbf{x})$ that guide transmitter localization. An Upper Confidence Bound acquisition $\mathrm{UCB}(\mathbf{x};\beta_n) = \mu(\mathbf{x}) + \beta_n \sigma(\mathbf{x})$ drives the next sampling location, balancing exploration and exploitation as the drone operates under a narrowband channel model. The evaluation spans simulation, lab emulation, and an AERPAW flight, demonstrating localization within tens of meters under favorable conditions while highlighting transfer gaps between emulation and real-world data that warrant a more tightly integrated digital twin for robust field deployment.

Abstract

This paper explores how a flying drone can autonomously navigate while constructing a narrowband radio map for signal localization. As flying drones become more ubiquitous, their wireless signals will necessitate new wireless technologies and algorithms to provide robust radio infrastructure while preserving radio spectrum usage. A potential solution for this spectrum-sharing localization challenge is to limit the bandwidth of any transmitter beacon. However, location signaling with a narrow bandwidth necessitates improving a wireless aerial system's ability to filter a noisy signal, estimate the transmitter's location, and self-pilot to improve the location estimate. By showing results through simulation, emulation, and a final drone flight experiment, this work provides an algorithm using a Gaussian process for radio signal estimation and Bayesian optimization for drone automatic guidance. This research supports advanced radio and aerial robotics applications in critical areas such as search-and-rescue, last-mile delivery, and large-scale platform digital twin development.

Figures (4)

• Figure 1: This figure shows the AERPAW boundaries for the AFAR challenge. The blue area indicates the aerial vehicle boundary limits. The green area indicates the possible locations of the rover. The numbered red points indicate where the rover was hidden, similar to Fig. 1 of kudybaUAVChallAERPAW. The white waypoints indicate the locations sampled within our algorithm's startup routine for each run.
• Figure 2: This figure shows the final results of a robotarium trial showing the GP mean (left) similar to Fig. 4a of kudybaUAVChallAERPAW, and the GP uncertainty as a mean squared error (right). Yellow is a high value, and blue is a low value for both images.
• Figure 3: This figure shows the output of the quality-variance filter along with the signal intensity given from the AERPAW narrowband channel sounder script. This data was collected in a lab setting that resembles the AERPAW setup without any RF front end. The receiver was placed on a moving platform, which was moved from 16 ft to 4 ft and then back to 16 ft from the transmitter. This figure is adapted from Fig. 4.2f of kudybaThesis.
• Figure 4: This figure shows the first and second run reconstructions of the final GP estimates over the search spaces shown in figure \ref{['fig/overview']}. The actual rover location is shown as a red dot. The green dot shows the final ten-minute estimate, and the blue dot indicates the three-minute estimate. This GP reconstruction uses the final kernel hyperparameters indicated by the vehicle log. Part of this figure is adapted from Fig. 4c of kudybaUAVChallAERPAW.