mems-nav-dataset — An open benchmark for MEMS-grade navigation
Existing navigation datasets tend to fall into two camps. The robotics and unmanned systems community publishes short, small-scale laboratory or field collections. The marine and aerospace communities publish long trajectories, but on high-end sensors most people will never touch. Neither is a good benchmark for the case that actually dominates in practice: long-duration navigation on cheap, noisy, MEMS-grade hardware.
This dataset splits the difference. It is collected on what is plausibly the most ubiquitous sensor configuration in existence — the MEMS IMU and GNSS receiver in a modern smartphone — over trajectories long enough to be comparable to marine and aerospace collections.
Raw collections include accelerometer, gyroscope, magnetometer, gravity, barometer, orientation, and WGS84 GNSS fixes, downsampled and time-synchronized. On top of those, the repository ships processed scenarios generated through strapdown-sim covering the failure modes that alternative-PNT research actually needs to test against:
| Scenario | What it models |
|---|---|
baseline | Standard closed-loop INS with GNSS updates, no degradation |
sched_10s, duty_10on_2of | Fixed-interval and duty-cycled GNSS availability |
degraded_fullrate, degraded_5s | Correlated measurement noise and inflated covariance |
slowbias, slowbias_rot | Slow drifting bias, with and without platform rotation |
hijack | Spoofing: a constant offset injected over a fixed interval |
combo, combo_duty_hijack | Combined degradation, intermittency, and spoofing |
The intent is that this is usable both as a teaching dataset and as a shared baseline, so that GNSS-denied and GNSS-degraded navigation results become comparable across papers.
Citation
@misc{brodovsky_mems_nav_dataset,
author = {Brodovsky, James},
title = ,
year = {2025},
howpublished = {\url{https://github.com/jbrodovsky/mems-nav-dataset}},
note = {Zenodo DOI forthcoming}
}