Velocity-consistency candidate v9 results: No-Go (6/558 PPC, 2/186 UrbanNav)¶
The frozen contract is online_pva_candidate_v9.md.
- Freeze: commit
c06567c1, together with the implementation, before any comparison replay. - Contract SHA256:
8313064b98440acbcd69e2d35923071da73785bc1925b6a8821969c452c4381d. - Default: unchanged.
- Tuning: nothing was tuned after the results were seen.
- Machine-readable record: online_pva_decision_v9.json.
Decision¶
No-Go. 6 of 558 PPC gates fail and 2 of 186 UrbanNav gates fail. This is the fewest failures of any candidate so far.
- Replays: 48 replays from one binary (SHA256
53135341..., built from the frozen worktree). They were interleaved, and all passed. - Gate 7: the control
noneis bit-identical to developc37979c3on the 18 PPC runs (175,902 rows). - Processor P95: candidate / control is 0.99 / 1.31 / 1.88 (min / mean / max).
| Run/scenario | Failed gate (both cohorts) | Control | Candidate |
|---|---|---|---|
| Tokyo 1 | rotation RMSE | 105.1 deg | 106.3 deg |
| Tokyo 2 | rotation RMSE | 100.3 deg | 103.9 deg |
| Tokyo 2 GNSS outage | rotation P95 | 68.1 deg | 73.4 deg |
| Odaiba IMU gap | RTK position P95 | 26.8 m | 28.5 m |
Normal runs: control / v7 / v8¶
RTK and fused values are RMSE / P95. Rotation is RMSE.
| Run | RTK pos m | RTK vel m/s | Fused pos m | Rotation deg |
|---|---|---|---|---|
| Tokyo 1 | 32.2/61.7 / 21.0/35.3 / 21.1/34.7 | 2.10/3.29 / 1.75/3.39 / 1.17/2.22 | 71.7/155.2 / 12.7/33.6 / 10.8/17.2 | 105.1 / 1.9 / 106.3 |
| Tokyo 2 | 19.5/29.7 / 6.3/10.6 / 6.3/10.6 | 2.26/3.71 / 1.08/2.20 / 0.75/1.59 | 74.2/179.5 / 4.2/8.8 / 3.6/7.5 | 100.3 / 1.4 / 103.9 |
| Tokyo 3 | 18.6/31.9 / 9.0/22.5 / 12.0/23.9 | 2.10/1.76 / 0.97/1.98 / 0.66/1.30 | 41.9/115.0 / 5.9/14.7 / 10.4/9.0 | 63.5 / 1.6 / 1.6 |
| Nagoya 1 | 25.7/29.0 / 21.2/27.5 / 20.9/26.7 | 2.51/4.26 / 2.02/4.55 / 1.28/2.36 | 49.3/118.1 / 14.1/25.1 / 12.5/11.3 | 88.2 / 2.2 / 1.8 |
| Nagoya 2 | 37.7/81.8 / 30.6/72.5 / 30.9/71.7 | 2.30/5.09 / 1.69/4.11 / 1.49/2.82 | 81.0/168.0 / 21.4/53.3 / 7.3/17.1 | 18.5 / 2.8 / 2.8 |
| Nagoya 3 | 43.7/101.1 / 36.0/99.2 / 37.0/99.5 | 2.47/5.13 / 1.62/3.78 / 1.50/3.31 | 58.9/128.1 / 34.1/94.6 / 14.2/19.9 | 105.0 / 1.9 / 1.5 |
| Odaiba | 199.9/47.8 / 199.1/27.9 / 199.0/28.3 | 3.84/9.31 / 2.20/4.96 / 2.32/5.00 | 81.2/157.1 / 12.5/31.7 / 19.6/57.2 | 105.6 / 104.4 / 105.6 |
| Shinjuku | 27.6/65.6 / 97.9/37.1 / 18.2/38.5 | 1.96/3.96 / 4.47/10.75 / 1.02/1.31 | 82.8/175.1 / 231.4/24.3 / 7.3/13.7 | 68.7 / 12.6 / 8.8 |
Reading (post-hoc; not used to change the candidate)¶
What worked¶
- Shinjuku is resolved. The 9.6 km base-seeded FLOAT is gone. RTK RMSE falls from 97.9 to 18.2 m, fused RMSE from 231 to 7.3 m, and rotation is 8.8 deg.
- RTK velocity. Every RTK velocity P95 gate passes. The epoch SPP velocity is better than the control's on every run.
- UrbanNav availability. It passes on every scenario, including Odaiba, against the expectation stated in the contract.
- Nagoya 2. The first heading latch matches the control.
- Fused position. It improves further on five of eight runs.
What regressed¶
- Tokyo 1 and Tokyo 2 attitude. It is lost, where v7 had 1.9 and 1.4 deg.
- How it fails on Tokyo 1:
- The heading error is -16.7 deg at 16 s, -44 deg at 20 s, +94 deg at 60 s and +161 deg at 120 s. It keeps wandering for the rest of the run.
- That is a drifting heading from the first latch on, not a 180 deg flip.
- Fused position and velocity stay good meanwhile: fused velocity RMSE 1.16 m/s.
- Likely cause. The only v8 change that touches the attitude path from the start is (n). With it, the epoch SPP velocity and its covariance become the fused and RTK-prior filters' velocity measurement. Isolating it needs a new diagnosis.
Remaining failure¶
- Odaiba IMU gap: RTK P95 is 6 % worse (26.8 -> 28.5 m).
Next¶
- A follow-up should isolate (n) on Tokyo 1 and Tokyo 2.
- It should then decide under a new frozen contract whether the SPP velocity should feed only the exported RTK velocity, which fixes the velocity gates, and not the fusion input.
- All eight runs are development data.