Photonic vs MEMS, proven on real roads

Case Study

A 21 minute drive through Montreal with GNSS positioning denied, comparing the MK2E2 optical gyro against the MK2M2 MEMS unit. Heading came from the gyro alone, with a 1 Hz speed input for distance and no EKF. Both units publish the same 0.5 °/hr minimum in-run bias. Only one was still close to truth at the end.

Results

  • 5.7 m final position error, MK2E2
  • 20.5 m final position error, MK2M2
  • 3.6x smaller final error

Pure Dead Reckoning: The route, with GNSS off.

Dead reckoning trajectory through Montreal with GNSS positioning denied, showing GNSS truth, MK2E2 photonic and MK2M2 MEMS paths.

With GNSS position removed, each unit had to keep track of where the vehicle was pointing using its gyro alone. Distance came from a 1 Hz speed input. Nothing corrected the heading once the drive began, so heading error carried forward with nothing to take it back out.

The difference shows up most clearly on the long straight segments. MEMS bias pushes the red path steadily sideways off the true route, and because heading error compounds with distance, each straight makes the gap worse. The photonic path tracks the true route through the same turns and the same traffic.

Both units sat in the same vehicle, on the same drive, logging at 100 Hz, on the same interface and the same supply. What changes between them is the Z-axis rotation sensor.

Error over 21 minutes: Drift, minute by minute.

Position error over the drive. The MK2M2 MEMS error rises above the MK2E2 photonic error from about the two minute mark and stays above it, ending at 20.5 m against 5.7 m.

This is the same drive measured as distance from truth, second by second.

For the first two minutes the two units are hard to separate, and the traces come close again around the four minute mark. After that the MEMS error runs consistently higher for the rest of the drive.

Neither error accumulates cleanly, and the plot is honest about that. Both traces rise and fall with the route. The MEMS trace sheds roughly 8 m at around fourteen minutes, and both units spike together near seventeen minutes, MEMS to about 36 m and photonic to about 22 m, before falling back. A spike common to both units points to the route or the shared inputs rather than to either gyro.

What persists is the separation between the two, not any single value. The photonic unit recovers to finish at 5.7 m. The MEMS unit settles at 20.5 m.

Where each ended up: Final position vs truth

Magnified view of the end points. The MK2E2 photonic finishes 5.7 m from truth, the MK2M2 MEMS 20.5 m.

At the end of the drive, with no position fix throughout, the MK2E2 was 5.7 m from where GNSS says the vehicle actually was. The MK2M2 was 20.5 m away, roughly four car lengths further from truth.

Both units finished on the same block, which is what a 15 m difference looks like at street scale. The distinction that matters is what you can do with the answer. At 5.7 m you are within a vehicle length of the true position, still on the correct street and the correct side of it. At 20.5 m you can be at the wrong building frontage, or on the wrong carriageway of a divided road.

That gap was produced by one component change, over 21 minutes, with no position aiding of any kind.

Method: How the test was run

The test was deliberately simple. No Kalman filter, no map matching, no clever post-processing. Just heading integrated from each gyro, combined with a speed input, so the comparison reflects the sensors rather than the software wrapped around them.

Setup

Montreal, 9 January 2026. A Dodge Journey driven through the downtown core for 21 minutes at speeds up to 50 km/h, covering ordinary city conditions: traffic, stop-and-go, turns, and the vibration of a normal road surface.

Both IMUs were mounted in the same vehicle for the same drive and logged at 100 Hz. GNSS was recorded throughout. Its position output was withheld from both solutions and used only as ground truth for scoring afterwards. The 1 Hz speed input was derived from the same GNSS log, so speed is the one aided quantity in this test. Heading was never aided.

Processing

Heading was obtained by integrating each unit’s Z-axis gyro. Position was propagated from that heading plus the 1 Hz speed input. No EKF, no GNSS position aiding, no zero-velocity updates, and no tuning specific to either unit. Both datasets went through the identical pipeline.

Error is reported as horizontal distance between the dead-reckoned position and the GNSS-recorded position at the same instant.

One note on scope. A production installation would take distance from wheel odometry and would run an EKF, and both would improve the absolute numbers. This test strips that layer out deliberately, so that what is being compared is the gyro.

What this means for your integration

Every inertial system drifts once the fix is gone. The question is how fast, and how long you can operate before the number stops being useful.

Talk to an engineer about your outage profile.