UAV navigation and flight continuity (R6)¶

Status: development navigation and visualization pass the frozen profile; mapping is No-Go. The untouched holdout run1 stopped fail-closed at the public MCAP derivative's invalid embedded schema encoding, so the R6 release decision is No-Go for navigation and visualization promotion. The failure is preserved without reader changes, threshold changes, or a second holdout run.

Application boundary¶

This route targets one MARS-LVIG flight container containing ZED-F9P raw GNSS, Livox IMU, DJI attitude, independent DJI RTK position/yaw, and fixed CAD antenna offsets. It produces an antenna-phase-centre standalone trajectory. It does not claim body-origin accuracy, centimetre accuracy, a production flight controller, or a mapping reference.

MARS-LVIG's official dataset page does not state a dataset licence. The article record is CC BY-NC-ND 4.0, which is not evidence that downloaded flight containers may be redistributed. Keep source data under ignored data/, fetch it for local evaluation, and do not ship it with libgnss++ artifacts.

The frozen profile is configs/benchmarks/uav_r6_mars_lvig_development.json. It records source sizes, SHA-256 values, topic counts, frame/time contracts, IGS broadcast NAV products, segmentation rules, and thresholds.

One-command route¶

After acquiring the named flight and uncompressed IGS mixed broadcast NAV, run:

python3 apps/gnss.py uav-mars-workflow \
  --input data/mars_lvig/cache/HKisland_GNSS03.bag \
  --navigation data/mars_lvig/navigation/BRDC00IGS_R_20232970000_01D_MN.rnx \
  --profile configs/benchmarks/uav_r6_mars_lvig_development.json \
  --role development \
  --output-dir output/uav_r6/development/reproduction

Use --max-epochs 600 for a bounded acquisition/adapter/native-SPP smoke. A bounded run deliberately does not spend a sign-off population decision because the required cruise, turn, climb/descent, hover, and landing populations may be incomplete.

The workflow verifies the container and NAV hashes before reading them, then runs the adapter, native gnss spp, frozen sign-off, KML, PNG, and hash manifest stages. Important outputs are:

uav_observations.rnx: mixed GPS/GLO/GAL/BDS RINEX 3.04 observations;
imu.csv, attitude.csv, rtk_truth.csv, rtk_yaw.csv, rtk_status.csv;
libgnsspp_spp.pos and libgnsspp_spp_summary.json;
uav_matches.csv and uav_signoff_summary.json;
libgnsspp_spp.kml, libgnsspp_spp_trajectory.png, and workflow_manifest.json.

The raw receiver PVT is used only for the solver's approximate initial ECEF position. DJI RTK fields are adapter outputs and scoring truth; they never seed or correct the native solution.

Frames, time, and lever arm¶

The actual bag declares Livox IMU frame livox_frame and DJI attitude frame body_FLU. DJI's ROS conversion defines the quaternion as body FLU to ground ENU. Attitude alignment therefore compares RTK yaw with (-yaw_ccw(body_FLU→ENU)) mod 360 and fits no offset.

DJI's RTK NavSatFix leaves frame_id empty. The adapter accepts that only because the frozen profile explicitly declares WGS84 latitude/longitude and ellipsoidal height at the DJI RTK antenna phase centre. Any other undefined frame fails.

CAD places the raw and truth antenna phase centres 0.615727 m apart in the LiDAR frame. The LiDAR-to-DJI-body fixed rotation is not authoritatively resolved, so the adapter does not rotate or apply that vector. Navigation errors are antenna-to-antenna comparisons with the 0.615727 m reference-point mismatch reported separately. Vehicle NHC is always disabled.

Sensor truth is UTC; raw observations preserve GPST week/TOW and use 18 leap seconds for 2023 comparison. Truth interpolation is allowed only across adjacent RTK samples no farther than 0.25 s apart.

Frozen development evidence¶

HKisland_GNSS03 used the official ROS1 bag SHA-256 8ce44d1d...c5dc317 and 2023 DOY 297 IGS NAV SHA-256 d5135d67...7966288. Native SPP produced 3,911/3,911 solutions.

Population	Epochs	Horizontal p95	Absolute vertical p95
Whole route	3,907	6.253 m	8.464 m
Cruise	2,030	6.362 m	8.941 m
Turn	552	6.005 m	8.954 m
Climb/descent	666	4.656 m	7.097 m
Hover	1,146	5.235 m	8.050 m
Landing tail	300	4.680 m	5.915 m

Availability was 100%, truth coverage 99.898%, maximum solution gap 0.100 s, maximum horizontal jump 2.135 m, attitude/RTK-yaw p95 6.322 degrees, and both RTK fixed-status ratios 100%. All 28 frozen gates passed. The immutable summary is docs/use_cases/records/uav_r6_development.json.

Development consumer decisions:

navigation: Pass for the named standalone antenna trajectory and thresholds;
visualization: Pass;
mapping: No-Go because independent truth is four-DoF and the fixed LiDAR/body rotation is unresolved.

Untouched holdout run1¶

Thresholds were committed at 0dfcc4e; the derivative container, NAV, and workflow were sealed at 5f14b31, 5e17ba2, and 604b3ce. HKairport_GNSS03 run1 used the published MCAP SHA-256 d85e39b9...7951de7. Acquisition passed, but the adapter stopped before deserializing messages because rosbags rejected an embedded MCAP schema with an empty encoding (KeyError: ''). No position, truth score, KML, or PNG was produced.

The failure is a container/provenance compatibility failure, not an accuracy result. It makes navigation, mapping, and visualization promotion No-Go. The immutable record is docs/use_cases/records/uav_r6_holdout_run1.json. Fixing the reader and trying the same holdout would tune on an opened holdout; a future R6 attempt needs a new versioned profile and a different sealed flight.

Reject the workflow when¶

source size/hash, NAV date/hash, or expected topic counts differ;
a quota/login page is returned instead of the frozen ROS1/MCAP magic;
time moves backwards, truth interpolation crosses a gap, or required motion populations are empty;
IMU/attitude frames are empty or differ from the frozen contract;
lever-arm values are missing/non-finite or vehicle NHC is enabled;
attitude quaternions are non-unit, RTK fixed status is insufficient, or any frozen coverage/error/gap/jump/yaw/landing gate fails;
a consumer asks for body-origin or mapping accuracy from this four-DoF truth.