Domain Realtime GPS Hardware

Domain -- The GPS hardware rig

The streaming tracker on real silicon and real fixes: an Arduino Nano 33 BLE Sense (nRF52840, Cortex-M4F) with a NEO-M8N GPS, logging raw fixes, the 9-axis IMU and an on-chip float32 estimate to an SD card and over BLE.

Notebook: rig in packages/dtfit-hardware. It reruns without the board: the serial captures of the on-chip runs and the compile sizes are tracked, and the notebook recomputes the float64 reference and compares. Flashing and capturing is a written procedure in the notebook. The recorded drives are not part of the repository; the notebook reports error tables only.

Routes and baselines

• dtfit: the GPS-only Legendre window tracker with the adaptive and the fixed window; LocalTimeFilter, the same tracker carried in the time of its newest fix; the on-chip filter (window 15, Legendre order 5, a two-parameter model per axis) in float32.
• Baselines: Kalman-CA and a CT-EKF on the same fixes.

No surveyed truth exists for the three logs (a 1 Hz drive, a 5 Hz drive, a walk): the scores are forecast against the later fix and coast against held-out fixes.

What is measured

Forecast. At a common configuration CT-EKF and Kalman-CA win the shortest horizons (two samples on the 5 Hz drive and the walk, three on the 5 Hz drive) and past that the tracker wins: 5.43 m against 6.18 and 8.62 m for the two baselines at ten samples on the 5 Hz drive, 19.21 m against 31.18 and 53.47 m on the walk. With each method's configuration chosen on the other two logs the tracker leads Kalman-CA on the 5 Hz drive and the walk everywhere but the walk's two-sample horizon, which is level; against CT-EKF it is 7 percent behind at two samples, level at three and ahead at ten.

The adaptive window wins both driving logs at every horizon and dropout; on the walk it wins the short horizon and loses the 10- and 15-sample dropouts to the fixed window (15.14 against 14.74 m, 21.83 against 21.13 m).

The clock origin. The absolute-time tracker is not blind to the clock's zero: shifting a log's time stamps by 10000 s moves its two-sample RMSE from 1.73 m to 8.91 m on the 5 Hz drive. Moving the polynomial's origin onto every new fix, exact for a polynomial, removes that: with the window held no shift moves a score. The re-centred tracker takes the 1 Hz drive from 5.85 m to 4.14 m at two samples, level with Kalman-CA and ahead of it from three; it is 1 to 7 percent behind the absolute-clock tracker on the 5 Hz drive and a fifth behind it on the walk.

On the chip. Across the three sketches flash use is 9 to 39 percent and RAM 17 to 28 percent of the board; the filter's struct is 152 bytes and an update 104 microseconds. Its float32 output agrees with a float64 replay of the same algebra to millimetres over a full drive.

Where dtfit loses

• On the 1 Hz drive CT-EKF and Kalman-CA beat the GPS-only tracker at every horizon with a common configuration; chosen per log the tracker's place against Kalman-CA starts about five samples out, and against CT-EKF the grid gives it none.
• Under dropouts CT-EKF coasts better than either window setting at every gap of both driving logs and at the walk's two shorter gaps.
• The metre-scale gap between the chip's output and the library's own filter is the damped-step guard and the jump test the chip does not run.

Three recorded drives, one board, one frozen on-chip configuration.