|
2 | 2 |
|
3 | 3 | ## Unreleased |
4 | 4 |
|
| 5 | +## 1.2.0 - 2026-07-22 |
| 6 | + |
| 7 | +Tightly coupled GLIL, exact point-cloud downsampling, and guarded kidnapped-pose |
| 8 | +recovery. Release boundary: PR #128 merged at `04bc8c7`; Humble and Jazzy CI passed, |
| 9 | +the Python suite reported 314 tests plus 140 subtests passed, and the pinned GLIL |
| 10 | +container passed its embedded CTest suite. |
| 11 | + |
| 12 | +### Tightly Coupled GLIL |
| 13 | + |
| 14 | +- Combined IMU preintegration, exact-coreset scan-to-scan GICP, and exact-coreset |
| 15 | + scan-to-prior-map GICP in one GLIM fixed-lag factor graph, following the estimator |
| 16 | + structure of Koide et al. ICRA 2024 and deferred exact downsampling of Koide et al. |
| 17 | + ICRA 2025. |
| 18 | +- Added continuous `odom -> base_link` output through map loss and one bounded |
| 19 | + `map -> odom` authority; map-factor loss, rejected candidates, and recovery never |
| 20 | + reset raw odometry or introduce a second TF parent. |
| 21 | +- Added verified map-state epochs, time-aligned 32-candidate BBS/GLIL re-ranking, and |
| 22 | + three-frame full-resolution consensus for kidnapped-pose recovery. |
| 23 | +- Recovery loss now checks both overlap fraction and the absolute map-factor inlier |
| 24 | + floor, closing the misleading 12/12-style sparse-overlap case while ignoring empty |
| 25 | + transport frames. |
| 26 | +- Published reproducible 01a/01b/02a/02b and outdoor-kidnap A/B runners, measurements, |
| 27 | + and six GLIL-only replay GIFs in `docs/koide_gif_gallery.md`. |
| 28 | + |
| 29 | +### Validation Boundary |
| 30 | + |
| 31 | +- Full 01a, 01b, and 02a runs passed accuracy, continuity, real-time playback, queue, |
| 32 | + and TF/reset gates. The final 02b run passed accuracy and TF gates; its CPU-contended |
| 33 | + playback result is explicitly labeled rather than claimed as clean throughput. |
| 34 | +- Outdoor kidnap A and B both ended `recovered_true`, at 0.074 m and 0.041 m final |
| 35 | + error respectively, with zero TF jumps, zero unauthorized resets, and drained final |
| 36 | + queues. |
| 37 | +- The tightly coupled GLIL path is the measured main configuration; NDT is not launched |
| 38 | + alongside it. Legacy split prior-map localization remains available for comparison. |
| 39 | + |
5 | 40 | ### Added |
6 | 41 |
|
7 | 42 | - `failure_category` key on `/alignment_status` distinguishing |
|
0 commit comments