Skip to content

technical_report_v0.3.0.md

Latest commit

 

History

History
103 lines (86 loc) · 5.17 KB

technical_report_v0.3.0.md

File metadata and controls

103 lines (86 loc) · 5.17 KB
Raw

gnss_gpu v0.3.0 technical report

Scope

v0.3 turns the urban-navigation research workspace into a guarded GNSS/IMU/map/GPU platform with immutable negative controls, truth-free acceptance evidence, multi-hypothesis outage recovery, enforced real-time budgets, cross-domain validation, and a ROS 2 lifecycle safety boundary.

Audited results

Contract Locked result
Historical negative controls 4/4 rejected
DDPR structure 2/3 WP163 ranks recovered; 0 WP164 false passes
Outage recovery 3 evidence epochs; greedy 10.0 m vs retained 0.0 m
GTX 1660 Ti runtime normal max 13.761 ms; search max 75.907 ms
WP172 candidate supply 42.653 ms/epoch conservative sequential; 21.826 ms/epoch parallel
Runtime capacity 152.588 MiB conservative estimate; 0 deadline misses
Cross-domain positioning epoch-weighted RMS 17.107 m to 16.916 m
Campaign coverage 3 cities; 9 sites/routes; 5 dates; 3 receivers
ROS replay 10 events; 1 restart; canonical event hash locked
ROS continuity 2 simulated hours; 439,187 events; 1 watchdog trip; final normal
Guarded MLAMBDA FIX Tokyo 10.8688%; Nagoya 18.0667%; false-FIX 0
Production promotion 12/12 gates pass; Tokyo 46.5112% exceeds 45%

The cross-domain result preserves Tokyo and improves the tracked Hong Kong campaign. Every release claim is derived from locked, hash-checked evidence included in the reproducibility archive.

Safety and promotion

Promotion is fail closed. It requires truth-free production input, positive full-denominator gain with zero loss, zero false FIX, all mandatory negative holdouts rejected or explicitly abstained, exact M4 preservation, multi-city non-degradation, reproducible input/config hashes, and the Phase 4 latency/memory limits.

WP172 seeds a demo5-continuous RTK candidate from the locked PF trajectory, then accepts positions only when an independent RTK trajectory agrees within 1 m and the prefit residual RMS is at most 4 m. The gate was frozen on Nagoya run1, where it gains 441 epochs with zero loss, before the Tokyo final audit. Tokyo gains 1,802 epochs with zero loss and reaches 5,546/11,924 sub-50 cm epochs (46.5112%). The selector is PF-only, uses no runtime FGO, and does not load reference truth until the post-selection full-denominator audit.

WP173 promotes the existing libgnss++ MLAMBDA result to a declared FIX only when the complete WP172 consensus gate passes, the LAMBDA ratio is at least 3.0, at least six satellites participate, and eligibility remains contiguous for five epochs. These standard ratio/hold conditions were frozen on Nagoya: 1,370/7,583 epochs (18.0667%) are declared FIX with zero false FIX. Applied unchanged to Tokyo, 1,296/11,924 epochs (10.8688%) are declared FIX with zero false FIX. A failed ratio, satellite count, continuity check, or missing candidate immediately falls back to FLOAT.

The two RTK candidate paths were replayed over all 11,924 Tokyo epochs on an i7-9750H. Their conservative sequential sum is 42.653 ms/epoch, below the 100 ms production bound; concurrent wall time is 21.826 ms/epoch. The seeded RTK output and final WP172 trajectory are byte-identical to the locks. The independent RTK path itself is not byte-identical, but the guarded final trajectory remains byte-identical, which is the product-level determinism boundary.

The ROS 2 lifecycle boundary rejects invalid parameters and timestamp regressions, publishes diagnostics, enters safe fallback under watchdog conditions, and produces a deterministic replay digest. A deterministic two-hour, 360,000-tick continuity audit injects duplicates, timestamp faults, a GNSS outage, and a restart; it records one watchdog trip, five recoveries, a bounded core state, and a final normal navigation mode.

Reproduction

Build and verify the release archive:

python tools/build_release_bundle.py \
  --output dist/reproducibility \
  --archive dist/gnss_gpu-v0.3.0-reproducibility.zip
python tools/build_release_bundle.py --verify dist/reproducibility

The archive contains the locked source evidence, a benchmark summary, ablation records, a failure gallery, this report, and a SHA-256 manifest. ZIP entry order and timestamps are deterministic.

Limitations

  • Tokyo clears the 45% positioning and 10% guarded FIX promotion floors, but most epochs intentionally remain FLOAT; the result does not claim broad RTK FIX availability.
  • Tokyo run1 is not a virgin scientific holdout: earlier campaign diagnostics had already inspected it. WP172's numerical 1 m/4 m gates were inherited from physical safety limits, frozen and checked on Nagoya before the locked Tokyo rerun; the result is therefore an operational promotion audit, not an unbiased estimate from a never-observed benchmark.
  • Phase 3 city-scale accuracy is supported indirectly by later campaigns; its locked outage-controller audit is synthetic.
  • Hong Kong is reproduced from a tracked result summary because raw data is not in this checkout.
  • The lifecycle core and replay contract are host-tested; rclpy and colcon integration are validated by the ROS container build.
  • The Windows Phase 4 memory value is a conservative capacity estimate because per-process nvidia-smi memory was unavailable.