Wireless communication networks aren’t just for transferring data—they can also be powerful sensors for detecting human presence. By closely examining the Received Signal Strength Indicator (RSSI) between radio-frequency (RF) nodes, researchers can identify subtle changes in signal strength caused by people walking nearby. Someone passing through a room may alter the signals in that area, triggering noticeable fluctuations in RSSI levels. Yet this same effect can arise when someone is moving farther away, such as in a hallway, causing fluctuations in a neighboring room. Distinguishing between nearby vs. distant movement is a key challenge in developing reliable motion-detection systems.
Within the EdgeAI project, scientists at Signify, Almende, and the Technical University of Eindhoven, are creating AI models that run directly on each RF node. These models must determine whether an RSSI fluctuation stems from a person in the immediate vicinity or from activity farther away. However, these nodes also operate as part of a larger network. How should each unit share information with its peers? Should they pass around only the parameters or “knowledge” gained from local data, or should they also broadcast actual predictions—essentially telling each other, “I think there’s motion right here”?
The setup at Almende in which RSSI data is being logged between RF nodes.
Figura 2: Syntzulu architecture overview
Why Communication Matters
As the number of intelligent devices grows, the strategies used for inter-device communication become increasingly important. In some settings, immediate collaboration can boost overall performance, while in others, too much back-and-forth might introduce unnecessary complexity. One concept that helps us decide which strategy is best is the notion of “relative error.” Simply put, when a model makes a prediction, we can measure how far off it is in more than one way. For example, consider a model that estimates a house’s value at $300,000 when the actual price turns out to be $400,000. If we measure the error relative to the actual price, we get (400,000 – 300,000) / 400,000 = 0.25. But if we measure it relative to the predicted price, the error is (400,000 – 300,000) / 300,000 ≈ 0.33. As we will see later the choice for relative error will determine the type of inter-device communication. Before reading further, you might want to give yourself a minute to think about what makes most sense to you: “error relative to the prediction” or “error relative to the reality”.
Backwards KL-Divergence (Variational Bayes)
In 2024, the EdgeAI project explored an approximation called the backwards Kullback–Leibler (KL) divergence, also known as variational Bayes. This technique effectively measures the difference between reality and predictions from the perspective of the prediction itself (“error relative to the prediction”). When working out the equation below one finds that AI models need to share their predictions with neighboring nodes. Over time, each node refines its guesses through “discussions” with peers, leading to a network-wide consensus. This approach can learn quite rapidly but may result in “groupthink,” where all nodes align on a shared prediction—even if it’s incorrect. Additionally, constantly transmitting predictions demands extra bandwidth, a potential drawback in large or resource-constrained environments.
P(θ│X)≈min┬(q(θ))∫▒〖dθ q(θ) log〖(q(θ))/(P(θ|X))〗 〗
The backward KL-divergence. X: sensor data,
θ:predicted quantities,P(θ│X):prediction including uncertainty.
Forward KL-Divergence (Advanced Density Filtering)
Given these disadvantages, attention has turned to forward KL-divergence (also called Advanced Density Filtering). This approach tracks the error relative to the actual outcomes. Mathematically one finds that in this approach AI models need to exchange only their underlying knowledge rather than their direct predictions. As they’re not constantly broadcasting guesses, the system is less likely to converge on a single flawed viewpoint, and the bandwidth requirements go down as well. Although it’s not yet fully understood how quickly these models learn, there’s reason to expect that once they do reach convergence, their predictions will be more robust.
P(θ│X)≈min┬(q(θ))∫▒〖dθ P(θ|X) log〖P(θ|X)/(q(θ))〗 〗
The forward KL-divergence. X: sensor data,
θ:predicted quantities,P(θ│X):prediction including uncertainty.
Real-World Trials in Rotterdam
Currently, the company Almende in Rotterdam (Netherlands) is gathering real-world data to put both methods through their paces. In practice, buildings often introduce complexities—such as subtle architectural features or unexpected sources of interference—that can make techniques that look good on paper stumble in real life. By comparing backwards KL-based models and forward KL-based models in these authentic settings, researchers aim to pinpoint the strengths, weaknesses, and best use cases for each. In scenarios where a human operator can step in and correct the system, a faster method with occasional groupthink might be acceptable. But for automated facilities or security systems that can’t easily be overridden, a slow-yet-steadfast approach may be preferable.
The office at Almende where RSSI data is at present being collected.
The Bigger Picture
This journey highlights a broader trend in AI: as we deploy networks of smart devices, the question of how they should collaborate becomes as critical as the question of how effectively each individual device can learn. Relative-error definitions, communication bandwidth limits, and even the possibility of groupthink are all factors that shape performance. EdgeAI’s experiments remind us that “simple” design choices—like how to measure an error—can cascade into large-scale repercussions.
Ultimately, motion sensing through wireless signals is more than detecting a person in a hallway or a room. It’s about figuring out how distributed AI units can work together efficiently while avoiding misleading consensus or excessive data exchange. Whether you favor “forwards KL-divergence” for its robustness or “backwards KL-divergence” for its speed, understanding these trade-offs helps craft better AI solutions suited to different real-world contexts. As the EdgeAI project continues to explore these questions, we can expect to gain insights that apply far beyond motion detection—delivering lessons on how to build networked intelligence that’s simultaneously resource-efficient, accurate, and adaptable.