Traditional indoor localization relies on exhaustive "fingerprinting"—manually measuring ambient radio signals (such as Wi-Fi RSSI or 5G Channel State Information) at thousands of known locations to build a reference map. This process is tedious, expensive, and fragile, as any change in the physical environment breaks the mapping.
However, a powerful geometric insight eliminates the need for manual calibration: high-dimensional radio signals naturally reside on a low-dimensional manifold that mirrors the physical geometry of the environment. Because radio waves propagate continuously through space, physical proximity implies signal similarity. By learning to untangle this high-dimensional signal manifold, we can discover relative spatial coordinates and track moving targets without any explicit location labels.
The Core Intuition: Radio Geometry as a Spatial Proxy
Why should raw, unlabeled radio data contain a map of physical space? Consider a mobile device moving through a building. As it moves, the high-dimensional vector of received signal strengths or channel impulse responses changes smoothly.
If you take two signal readings at positions separated by only a few centimeters, the multipath reflections, fading patterns, and amplitudes will be highly correlated. If you move to the opposite side of the building, the signals will be completely different.
As formalized by Christoph Studer et al. in their seminal framework on "Channel Charting" (Channel Charting: Locating Users within the Radio Environment using Channel State Information), the high-dimensional torrent of data captured by modern multi-antenna base stations is simply a non-linear, distorted wrapping of a low-dimensional physical space. Because the underlying cause of signal variation is physical movement, the intrinsic dimensionality of the signal space matches the dimensionality of the physical space (typically 2D or 3D). Learning the shape of this radio data manifold allows network operators to recover the relative spatial relationships of users in a completely self-supervised manner.
[Physical Space (2D/3D)] ---> Continuous Propagation ---> [High-Dim Signal Space]
^ |
| v
True Coordinates <------- Manifold Learning/Unfolding <------- Latent Embedding
Unfolding the Latent Space
To mathematically extract physical coordinates from raw signal vectors, the problem can be framed as unsupervised dimensionality reduction or latent variable modeling. Here, physical coordinates are treated as hidden (latent) variables that generate the observed high-dimensional signals.
In the pioneering "WiFi-SLAM" framework, Brian Ferris et al. utilized Gaussian Process Latent Variable Models (GP-LVM) to optimize the latent 2D coordinates of a mobile device based entirely on raw, unlabeled Wi-Fi signal strengths (WiFi-SLAM Using Gaussian Process Latent Variable Models). The optimized latent space inherently reflects the physical geometry and topological connectivity of the building.
To prevent these generative models from creating redundant, overlapping maps that miss key physical features like sharp corners, Dacheng Tao et al. introduced a diversified generative latent variable model (A Diversified Generative Latent Variable Model for WiFi-SLAM). By forcing the latent map points to spread out evenly, the model preserves correct topological structures.
In modern cellular networks, deep learning architectures like autoencoders are widely used to compress high-dimensional Channel State Information (CSI) into low-dimensional representations. For instance, Qianqian Zhang et al. demonstrated that convolutional autoencoders can compress millimeter-wave (mmWave) channel data into a 3D radio-geometry map, ensuring that devices physically close to each other maintain proximity within the digital representation (Semi-Supervised Learning for Channel Charting-Aided IoT Localization in Millimeter Wave Networks).
Preserving Locality and Neighborhoods
The mathematical engine driving this manifold unfolding is locality preservation. If two signal profiles are close in the high-dimensional space, the algorithm must map them to nearby coordinates in the lower-dimensional space.
Taotao Huang et al. showed that explicitly maintaining the topological relationship between the signal space and physical space via locality-preserving canonical correlation analysis dramatically reduces spatial warping errors caused by environmental noise (Locality Preserving Canonical Correlation Analysis Distributed Localization Algorithm for Wireless Sensor Networks).
This neighborhood structure is robust enough to enable real-time tracking and network optimization without explicit geographic coordinates:
- Online Co-Localization: Jeffrey Junfeng Pan et al. developed an incremental graph-based manifold learning approach to simultaneously track mobile devices and access points in real-time, eliminating static offline fingerprinting (Online Co-Localization in Indoor Wireless Networks by Dimension Reduction).
- Dynamic Adaptation: To handle environments that change over time, Sinno Jialin Pan et al. introduced LeManCoR, which treats signal data from different time periods as multiple views of the same physical space, adapting old signal maps to new conditions (Adaptive Localization in a Dynamic WiFi Environment through Multi-view Learning).
- Cross-Space Transfer: Sinno Jialin Pan et al. further demonstrated that these learned manifold structures can be transferred across different areas of a building, minimizing the calibration effort required when expanding a network (Transferring Localization Models across Space).
- Vehicular Networks: Hanan Al-Tous et al. applied unsupervised manifold learning to preserve the neighborhood structure of radio signals in vehicle-to-vehicle (V2V) communications, creating logical maps that accurately predict link quality without GPS overhead (Multipoint channel charting-based radio resource management for V2V communications).
To scale these concepts to continuous data streams, Sueda Taner et al. designed a memory-efficient framework using Siamese neural networks (Channel Charting for Streaming CSI Data). By training on a tiny, curated "core" subset of signals that preserves local geometric relationships, the network instantly maps incoming real-time radio fingerprints to their relative spatial positions.
Grounding to Real-World Coordinates
Purely unsupervised manifold learning recovers the topology (the relative layout and connectivity) of an environment, resulting in arbitrary "pseudo-coordinates." To transform these relative charts into absolute real-world coordinates (meters or latitude/longitude), researchers introduce weak supervision or physical constraints.
Hybrid Semi-Supervised Learning
Instead of dense fingerprinting, models can use a vast pool of unlabeled signal readings to map the environment's global structure, combined with a tiny handful of labeled anchor points to scale and rotate the map. Hong Chang et al. utilized a semi-supervised manifold regularization framework (LeMan) to interpolate accurately between sparse labeled points by leveraging the underlying geometry of the unlabeled data (A Manifold Regularization Approach to Calibration Reduction for Sensor-Network Based Tracking). Similarly, Sebastian Thrun et al. constrained signal distributions to a one-dimensional manifold representing the physical layout of corridors, using the Expectation-Maximization (EM) algorithm to learn person-specific motion models from unlabeled traces (A Learning Algorithm for Localizing People Based on Wireless Signal Strength that Uses Labeled and Unlabeled Data).
Self-Supervised Geometric Losses
Rather than relying on manual labels, systems can exploit known hardware layouts. Sueda Taner et al. introduced a "bilateration loss" that combines known base station locations with self-supervised local constraints, allowing the network to output absolute real-world coordinates instead of arbitrary pseudo-coordinates (Channel Charting in Real-World Coordinates).
Auxiliary Sensor Fusion
Another powerful strategy leverages dead-reckoning data from Inertial Measurement Units (IMUs), such as pedometers or velocity sensors, to provide physical distance constraints over short time windows.
[High-Dim Radio Data] --------> Manifold Embedding Space
|
Enforce Distance
Consistency
|
[Auxiliary IMU Data] --------> Physical Distance Delta
Caifa Zhou et al. demonstrated that correlating crowd-sourced radio signal changes with IMU data allows a system to learn the exact relationship between signal similarity and physical distance in a self-supervised manner, aligning random pedestrian trajectories into a coherent reference map (Mining geometric constraints from crowd-sourced radio signals and its application to indoor positioning).
This concept also enables device-free tracking. In "Passive Channel Charting," Raffael Poeggel et al. used relative velocity data from a wearable sensor to supervise the learning of a low-dimensional spatial manifold from the channel impulse responses of a static Ultra-Wideband (UWB) mesh network (Passive Channel Charting: Locating Passive Targets using a UWB Mesh). Moving bodies create smooth, learnable perturbations in the environment's radio field; by forcing the latent embedding distances to match the physical movement distances, the network organizes raw radio reflections into a highly accurate tracking map.
The Modality-Agnostic Nature of Spatial Manifolds
Perhaps the most profound realization in this research direction is that the underlying spatial manifold is entirely independent of the sensor type used to observe it. Whether an observer captures Wi-Fi signal strengths, 5G CSI, UWB impulse responses, or camera images, the intrinsic geometry of the data manifold is governed by the exact same physical floor plan.
As shown by Farhad G. Zanjani, the same physical geometry can be recovered from either visual data or Wi-Fi signal patterns using weakly-supervised deep learning frameworks like OT-Isomap and WiCluster (Deep Learning Frameworks for Weakly-Supervised Indoor Localization).
By combining manifold learning with optimal transport, Farhad Ghazvinian Zanjani et al. created a generalized framework that treats localization purely as a geometric mapping problem between a high-dimensional data manifold and a topological map (Modality-Agnostic Topology Aware Localization). Because it bypasses modality-specific physics models or specialized feature extractors, this mathematical framework achieves decimeter-level accuracy across completely diverse sensor types, proving that the geometry of space itself is the ultimate regularizer.
Key Foundational Work
A Manifold Regularization Approach to Calibration Reduction for Sensor-Network Based Tracking (2006) — contributed 1 concept elements to this direction.
A Learning Algorithm for Localizing People Based on Wireless Signal Strength that Uses Labeled and Unlabeled Data (2003) — contributed 1 concept elements to this direction.