LoRaCompass: Robust Reinforcement Learning to Efficiently Search for a LoRa Tag

TL;DR

LoRaCompass tackles the problem of locating a LoRa tag from RSSI readings with a mobile sensor in large, varied environments. It introduces a robust exploitation mechanism built on a receptive-field RSSI feature extractor and a policy distillation loss, together with a closed-form, UCB-inspired exploration function to reduce decision uncertainty. The approach is trained in a realistic simulator and validated on ground and drone platforms across over $80 km^2$ of unseen environments, achieving a success rate above $90%$ within $100 m$ and near-linear search efficiency compared to baselines. The work demonstrates practical deployability with a single training site and shows potential for multi-sensor collaboration to further boost search performance.

Abstract

The Long-Range (LoRa) protocol, known for its extensive range and low power, has increasingly been adopted in tags worn by mentally incapacitated persons (MIPs) and others at risk of going missing. We study the sequential decision-making process for a mobile sensor to locate a periodically broadcasting LoRa tag with the fewest moves (hops) in general, unknown environments, guided by the received signal strength indicator (RSSI). While existing methods leverage reinforcement learning for search, they remain vulnerable to domain shift and signal fluctuation, resulting in cascading decision errors that culminate in substantial localization inaccuracies. To bridge this gap, we propose LoRaCompass, a reinforcement learning model designed to achieve robust and efficient search for a LoRa tag. For exploitation under domain shift and signal fluctuation, LoRaCompass learns a robust spatial representation from RSSI to maximize the probability of moving closer to a tag, via a spatially-aware feature extractor and a policy distillation loss function. It further introduces an exploration function inspired by the upper confidence bound (UCB) that guides the sensor toward the tag with increasing confidence. We have validated LoRaCompass in ground-based and drone-assisted scenarios within diverse unseen environments covering an area of over 80km^2. It has demonstrated high success rate (>90%) in locating the tag within 100m proximity (a 40% improvement over existing methods) and high efficiency with a search path length (in hops) that scales linearly with the initial distance.

Figures (23)

• Figure 1: Search path from an agent at initial location $X$ to a LoRa tag at $T$, with the RSSI heatmap shown as background. (a) Traditional exhaustive search, which is independent of RSSI, resulting in an inefficient path. (b) State-of-the-art (SOTA) approach soorki2024catch, achieving a shorter path under the same RSSI heatmap. (c) Our LoRaCompass, demonstrating a more efficient (shorter) search path for the same heatmap. (d) Search path under domain shift and signal fluctuation. The SOTA approach (purple line) fails to converge to the tag owing to path looping, whereas LoRaCompass (red line) remains efficient and robust, successfully converging to the tag.
• Figure 2: Illustration of our realistic simulator used for on-policy training. The observation function is simulated by RSSI histograms collected from each grid point.
• Figure 3: Distribution of LoRa gateway in an urban region of $40\hbox{km}^2$. Gateways are marked by blue labels.
• Figure 4: System diagram of LoRaCompass.
• Figure 5: Observation behind feature extraction: Within a sufficiently large receptive field (centered at the agent location), the RSSI tends to be higher closer to the tag, and lower further away from it. This pattern sheds light on tag direction independent of domain shift and signal fluctuation.