1 00:00:00,000 --> 00:00:05,003 This is a record of a verification pass over LabeloxAV, run against the live system. 2 00:00:05,003 --> 00:00:06,329 Three things were done. 3 00:00:06,329 --> 00:00:15,672 More real data was downloaded and ingested, every readable interface was called to see what answers, and the HD map was taken end to end for the first time. 4 00:00:15,672 --> 00:00:18,867 Every number spoken here came from one of those runs. 5 00:00:18,867 --> 00:00:30,101 Two KITTI raw drives were downloaded from the public archive and imported, adding four hundred and eighty five frames that each carry a synchronised Velodyne scan, a GPS fix and the drive's own calibration. 6 00:00:30,101 --> 00:00:35,800 A third drive stopped part way through its download and was left out rather than imported half complete. 7 00:00:35,800 --> 00:00:43,489 The corpus now holds 1,049 point clouds and about 159 million real laser points, against eighteen million before this pass. 8 00:00:43,489 --> 00:00:47,396 Measured ego poses went from one hundred and fifty four to 639. 9 00:00:47,396 --> 00:00:55,400 Everything else in the corpus is still estimated from single camera depth, and the tooling says so on every page that reads it. 10 00:00:55,400 --> 00:01:00,530 This is a drive in the session inspector, with the camera, the laser and the GPS on one clock. 11 00:01:00,530 --> 00:01:09,467 The point cloud here is measured by a sixty four beam scanner, not inferred from pixels, which is what makes the three dimensional work on this session checkable. 12 00:01:09,467 --> 00:01:13,594 Before this pass no session had both lane annotations and a position fix. 13 00:01:13,594 --> 00:01:21,619 One hundred and fifty seven frames of forty two thousand carried a fix, and not one of them had a lane, so the map could not be built at all. 14 00:01:21,619 --> 00:01:29,643 Lane proposal turned out to run locally rather than only on a rented GPU, so it was run across a drive that already had a fix on every frame. 15 00:01:29,643 --> 00:01:37,267 One hundred and ninety nine lanes on one hundred and fifty four frames, with forty one rejected for sitting off the drivable surface. 16 00:01:37,267 --> 00:01:40,201 Those lanes were then georeferenced into world space. 17 00:01:40,201 --> 00:01:46,916 One hundred and ninety nine map elements, every one placed using the drive's own calibration rather than a nominal lens. 18 00:01:46,916 --> 00:01:48,721 That distinction is not cosmetic. 19 00:01:48,721 --> 00:01:56,000 With the wrong lens this projection put a lane six metres ahead at fifteen metres, and one at seventy metres at over a kilometre. 20 00:01:56,000 --> 00:02:04,217 Fusion merged those into fifty three continuous boundaries, and the topology layer paired twenty nine of them into lanes with three links between them. 21 00:02:04,217 --> 00:02:08,655 The exports now carry twenty nine lanelet relations and fifty eight driving lanes. 22 00:02:08,655 --> 00:02:14,133 Both formats previously carried zero of each, which meant neither was a map anything could route on. 23 00:02:14,133 --> 00:02:17,851 Every readable interface was then called against the live database. 24 00:02:17,851 --> 00:02:19,823 Three hundred and fifty two of them. 25 00:02:19,823 --> 00:02:30,246 Two hundred and twenty answered, fifty three refused properly, forty four could not be reached because this corpus has no identifier to fill the address, and one returned a server error. 26 00:02:30,246 --> 00:02:38,133 The four hundred routes that change data were not called, because firing those blindly would train models, launch paid jobs and delete work. 27 00:02:38,133 --> 00:02:45,271 The failure was the annotator branch listing, which returned a bare server error whenever the versioning service was not running. 28 00:02:45,271 --> 00:02:46,889 That service is optional here. 29 00:02:46,889 --> 00:02:51,406 Assignments, tasks and merge requests are all in the database and work without it. 30 00:02:51,406 --> 00:02:58,600 It now refuses with the reason and says what is unaffected, which is what every other optional path in this codebase already did. 31 00:02:58,600 --> 00:03:07,914 Three thousand five hundred and six tests pass with nothing failing, seven hundred and three in the web suite, the typecheck is clean and the import contract holds. 32 00:03:07,914 --> 00:03:11,400 Seventy two of the seventy three pages load without a problem. 33 00:03:11,400 --> 00:03:13,800 The one not checked is the sign in screen. 34 00:03:13,800 --> 00:03:17,678 Two things in the map are not yet trustworthy and are worth saying plainly. 35 00:03:17,678 --> 00:03:24,334 Fusion produced one boundary eight and a half kilometres long on a drive of a few hundred metres, which is a merge running away. 36 00:03:24,334 --> 00:03:34,658 And the median paired lane came out two and a half metres wide where a German lane is three to three and a half, which says the pairing is still matching things that are not the two sides of a lane. 37 00:03:34,658 --> 00:03:35,968 The layer runs end to end. 38 00:03:35,968 --> 00:03:41,733 The geometry underneath it needs work, and calling that finished would be the easiest thing to get wrong here. 39 00:03:41,733 --> 00:03:43,157 So the honest summary. 40 00:03:43,157 --> 00:03:56,104 More real data is in, every readable route has been called and one genuine failure was found and fixed, and the HD map runs from a camera frame through to a routable export for the first time. 41 00:03:56,104 --> 00:04:00,849 The human verdict count has not moved: 788 of 600,133 labels carry one. 42 00:04:00,849 --> 00:04:05,933 That remains the thing most worth fixing, and no amount of code changes it.