Radio-based indoor and outdoor localization systems—leveraging Wi-Fi, 5G, Bluetooth, or UWB—traditionally rely on a simple premise: an environment possesses a unique wireless "fingerprint." By conducting a site survey to map physical coordinates to specific signal properties like Received Signal Strength Indicators (RSSI) or Channel State Information (CSI), a model can later infer a user's location by matching their real-time signal profile to this reference database.
The fatal flaw in this premise is the assumption that the physical world is static. In reality, environments experience constant temporal signal drift. As noted by Deng et al., physical changes such as moving furniture or relocating signal sources degrade positioning accuracy over time. Human traffic, moving machinery in a warehouse, and even atmospheric or weather variations continuously alter how radio waves propagate through a space, as observed by Wen et al.. When the physical layout shifts, the pre-recorded radio map becomes obsolete, necessitating frequent and labor-intensive manual recalibration that Yang notes is impractical for long-term deployments.
Why Signal Drift Breaks the Mapping
To understand why temporal drift degrades localization, we must look at how radio waves interact with physical space. A wireless signal does not travel along a single straight path; it bounces off walls, doors, and objects, creating a complex web of reflections known as multipath propagation.
When a piece of machinery moves or a door opens, it alters these reflection paths, as explained by Zhao et al.. This introduces unpredictable constructive or destructive interference into the fine-grained CSI or RSSI values captured by a receiver.
Static Environment:
[AP] ------> (Reflection Path) ------> [Object] ------> [Receiver] (Stable Fingerprint)
Dynamic Environment:
[AP] ------> (Blocked/Altered Path) --> [Moved Object] [Receiver] (Signal Drift)
Furthermore, the underlying network infrastructure itself is non-stationary. Access points are routinely moved, reconfigured, or decommissioned over time, a problem highlighted by both Adams et al. and Shokry et al.. If a localization model maps a coordinate based on absolute signal values from specific beacons, the removal or relocation of even a single beacon warps the latent spatial representation, leading to severe positioning errors.
Invariance Through Differential Features
One powerful intuition for bypassing signal drift is to stop relying on absolute signal values entirely. If an environmental shift or a change in user hardware occurs, absolute RSSI values fluctuate wildly. However, the relative difference between the signal strengths of two nearby access points often remains remarkably stable.
This is the core mechanism behind the differential radio map framework introduced by Wang et al.. Instead of storing absolute values, the radio map records the mathematical difference between pairs of access points:
$$\Delta \text{RSSI} = \text{RSSI}_{\text{AP}_1} - \text{RSSI}_{\text{AP}_2}$$
If an environmental change (such as a surge in human density) introduces attenuation, it tends to affect nearby signal paths proportionally. If both $\text{RSSI}_{\text{AP}_1}$ and $\text{RSSI}_{\text{AP}_2}$ drop by $3\text{ dB}$, the subtraction cancels out the environmental bias:
$$(\text{RSSI}_{\text{AP}_1} - 3) - (\text{RSSI}_{\text{AP}_2} - 3) = \text{RSSI}_{\text{AP}_1} - \text{RSSI}_{\text{AP}_2}$$
By shifting the model's inputs from absolute signal space to a relative differential space, the system isolates the spatial signature from transient environmental noise, maintaining tracking accuracy without requiring a map overhaul.
Continual Adaptation and Self-Evolving Maps
While relative features mitigate minor fluctuations, long-term structural changes—such as a permanent building remodel or a new warehouse layout—require the localization model itself to adapt dynamically. Rather than wiping out the model and starting over, researchers use domain incremental learning and self-supervised map evolution to track environmental drift.
Instead of relying on professional surveyors to rebuild the database, systems can continuously update their radio maps using live, "messy" crowdsourced data from everyday users. For example, Zhou et al. demonstrated that by adaptively learning the mathematical relationship between radio signal similarity and physical distance from random pedestrian trajectories, a reference map can align and evolve automatically over time.
[Crowdsourced Trajectories]
│
▼
[Self-Supervised Alignment] ──> [Continuous Map Evolution] ──> [Updated Radio Map]
▲
│
[Real-Time Signal Drift]
To make this continuous update process computationally viable and prevent the model from forgetting past layouts (catastrophic forgetting), frameworks must be highly selective about how they learn. The 5G-DIL framework proposed by Raichur et al. uses similarity-aware sampling based on Chebyshev distance to update models using as few as 50 data samples, drastically cutting retraining overhead.
Similarly, the ConSense framework developed by Deng et al. isolates and updates only the specific neurons that display instability when exposed to new signal environments, while expanding the model with lightweight attention parameters. This targeted adaptation allows localization frameworks to gracefully track concept drift in highly dynamic environments, ensuring long-term reliability without requiring massive data storage or constant human intervention.