FERMI: Flexible Radio Mapping with a Hybrid Propagation Model and Scalable Autonomous Data Collection

TL;DR

FERMI addresses the challenge of accurate radio mapping in occlusion-rich, large environments by decomposing propagation into direct-path physics and a learned multipath component. It introduces a hybrid propagation model paired with a scalable, multi-robot data-collection planner, enabling autonomous data gathering with minimal manual intervention. The approach demonstrates strong generalization to unseen Tx-Rx pairs, data-efficiency with sparse training data, and scalable planning that reduces transition costs as robot team size grows, validated in both simulation and real-world tests. The framework provides a practical tool for communication-aware robot collaboration and is open-sourced for community use.

Abstract

Communication is fundamental for multi-robot collaboration, with accurate radio mapping playing a crucial role in predicting signal strength between robots. However, modeling radio signal propagation in large and occluded environments is challenging due to complex interactions between signals and obstacles. Existing methods face two key limitations: they struggle to predict signal strength for transmitter-receiver pairs not present in the training set, while also requiring extensive manual data collection for modeling, making them impractical for large, obstacle-rich scenarios. To overcome these limitations, we propose FERMI, a flexible radio mapping framework. FERMI combines physics-based modeling of direct signal paths with a neural network to capture environmental interactions with radio signals. This hybrid model learns radio signal propagation more efficiently, requiring only sparse training data. Additionally, FERMI introduces a scalable planning method for autonomous data collection using a multi-robot team. By increasing parallelism in data collection and minimizing robot travel costs between regions, overall data collection efficiency is significantly improved. Experiments in both simulation and real-world scenarios demonstrate that FERMI enables accurate signal prediction and generalizes well to unseen positions in complex environments. It also supports fully autonomous data collection and scales to different team sizes, offering a flexible solution for creating radio maps. Our code is open-sourced at https://github.com/ymLuo1214/Flexible-Radio-Mapping.

Figures (11)

• Figure 1: Flexible Radio Mapping (FERMI) is a data-driven, accurate, and efficient framework for signal data collection and propagation modeling. The framework enables the prediction of signal strength between any two positions using only sparse data. We demonstrate radio mapping results in an indoor corridor environment (94m × 64m × 3m), where (a) illustrates the point cloud map of the scene, (b) shows a photograph of the environment alongside our robotic platform, and below compares signal data generated by different methods when a Wi-Fi transmitter is placed at the marked position. Ground truth data is collected through dense coverage by the robots.
• Figure 2: Overview of FERMI: (a) With the collected signal data, we first sample Line-of-Sight (LOS) points for both the transmitter (Tx) and receiver (Rx). Based on a physical model, we compute the direct path propagation from the Tx to its surrounding LOS points. Next, a neural network is employed to learn the signal strength generated by multipath propagation between each pair of LOS points from the Tx to the Rx, as well as the attenuation parameters of direct path propagation from the Rx LOS points to the Rx itself. Finally, the complex-valued signals from the Rx LOS points to the Rx are aggregated. The resulting values are compared with the signal strength data collected by the robots to compute the loss, which is used to train the multipath propagation network. (b) The autonomous data collection begins by dividing the scene into multiple regions that are mutually non-visible. A team of robots equipped with Wi-Fi access nodes departs from their initial positions, collaboratively planning to execute the most parallel and cost-effective transition sequences to traverse different regions and collect signal data.
• Figure 3: An illustration demonstrates the rendering process for multipath propagation. Each LOS point of the Tx acts as a retransmitter. The energy propagated from the Tx to a LOS point via direct path propagation determines the weight assigned to that point when combining signals at the corresponding LOS point of the Rx. After calculating the signals transmitted through multipath propagation between each pair of Tx and Rx LOS points via the network, the received signals at each Rx LOS point are combined with the weights of the corresponding Tx LOS points. The signal contribution of each ray is then computed by integrating along its path. Finally, the signals from all rays are summed to determine the total received signal through multipath propagation.
• Figure 4: An illustration showing the robot team configuration and how the corresponding collection state matrix elements are updated.
• Figure 5: A schematic for our visibility-based scene partitioning method. (a) After segmenting the scene's surface and obtaining representative points, we cluster each representative surface point. A point is added to a cluster if it has visibility to all surface points already in that cluster. If it lacks visibility to any point in the cluster, it will not be added. (b) Illustration of the final partitioning results.