A Transformer That Fuses WiFi and Motion Data Without a Shared Clock

The timing problem

Indoor positioning cannot usually rely on GPS, whose signals are weak inside buildings. WiFi fingerprinting estimates location by comparing a scan’s signal strengths with patterns associated with known places. An inertial measurement unit (IMU) tracks motion, giving a denser account of how someone or something has moved. Used together, the sensors may compensate for one another: WiFi can provide an absolute reference, while motion can help fill in between scans.

But the readings do not arrive in lockstep. A common approach is to resample them onto a shared timeline, or to process each sensor in a separate model branch and combine the results. Resampling can require padding or discarding information. Systems must also contend with readings that are unavailable or too old to be useful.

The authors’ alternative is a continuous-time set transformer. It represents each available observation as a token containing an encoded measurement, the sensor type and the time elapsed since acquisition. The model uses timestamps to calculate that elapsed time, but does not resample the streams onto a common clock. Its attention mechanism combines tokens across sensors and times. Because it treats them as a set, input order does not determine the result. The architecture uses modality-specific encoders to make tokens, then feeds them into a shared six-layer transformer fusion trunk—not a single attention layer—to combine information across sensors and times.

During training, the model is exposed to randomly dropped sensor types and individual observations. This is intended to help it operate when inputs are missing or outdated; it does not guarantee reliable performance in every situation.

How it was tested

The paper evaluates five datasets with different roles: a Webots simulation, two real datasets containing WiFi and IMU information, and separate WiFi-only and IMU-only benchmarks for testing the individual encoders. The main real-world test uses the MSILN site1/B1 collection, with training, validation and test paths recorded on separate days. This cross-session split tests performance on sessions not seen during training.

The authors compare the fused system with named reference methods, including WiFi fingerprinting, inertial dead reckoning and a learned WiFi–IMU fusion model. They also test missing and aged observations in simulation, where those conditions can be controlled. These experiments address different questions: some evaluate an encoder alone, some test sensor fusion, and the simulation offers controlled robustness checks.

What the results show—and don’t

On the held-out MSILN test session, the proposed model’s mean position error was 10.90 metres. The reported errors were 28.31 metres for the WiFi fingerprinting reference, 12.49 metres for dead reckoning from a starting position, and 52.69 metres for the learned fusion reference. On the validation session, however, its result was 16.67 metres, close to dead reckoning’s 16.88 metres. The authors caution that the test result should be read alongside validation: one WiFi-dense path accounts for about 28% of test samples, so the test figure may not represent a broad set of equally weighted situations.

The individual encoders had mixed results. The WiFi encoder beat the paper’s fingerprinting reference on the UJIIndoorLoc validation benchmark. The much smaller IMU encoder did worse than a larger inertial model on the RoNIN benchmark. In the controlled Webots simulation, the fused system achieved a 0.61-metre mean error, but its WiFi field was synthesized. The authors describe this result as optimistic relative to real deployment, not as a real-world accuracy estimate.

In simulation, errors rose gradually as WiFi readings became older, rather than increasing abruptly. Removing a sensor also showed that WiFi supplied an important absolute reference, while IMU alone accumulated more error. These experiments indicate how the system behaved under the tested simulated conditions; they do not establish performance across buildings, devices or real-world patterns of missing data.

A promising proposal, not a settled result

The study presents a way to fuse asynchronous sensor data and reports encouraging results against selected baselines on a real cross-session dataset. It does not establish that the approach will generalize to all buildings, devices or environmental changes. The cross-session evidence is limited, the test-set composition affects the headline comparison, and the inertial encoder underperformed a larger reference on a separate benchmark. The authors also note the limitations of the synthetic simulation and suggest exploring additional sensors and physical constraints.

This article is based on the supplied paper text and abstract for arXiv preprint 2611.13057v1. A preprint is not, by itself, evidence of peer review, and the supplied material does not establish whether this version has been reviewed or independently replicated. The results should be understood as findings reported by one study, not as a settled benchmark for indoor localization.