[sunzenshen]

ntre_skyline_ctg: navmesh rework

One file, neo/maps/nav/ntre_skyline_ctg.nav.

Dependencies and status

Needs: none. Measured on neo master 092203f36 with no open PR (A/B below). The bot changes this mesh carries data for are
in master now:

• NAV_MESH_AVOID annotation penalizes bot path costs neo#2170 ("NAV_MESH_AVOID annotation penalizes bot path costs", merged): AVOID path cost. The two areas at ladder 1's foot are NAV_MESH_AVOID.
• Bots hold an exact move direction on PRECISE areas neo#2175 ("Bots hold an exact move direction on PRECISE areas", merged): PRECISE. 37 areas are NAV_MESH_PRECISE: 35 rooftop fall edges (also CLIFF) and
  two at the stair corner near (-2265, -1900). On neo master, fall deaths from these 37 areas: 1.06 a match on master's mesh, 0.18 on this one.
• More tweaks to bot ladder approaches, climbs and dismounts neo#2176 ("Fix bot ladder approaches, climbs and dismounts", merged): ladder climbing. The mesh has 3 nav ladders, as master.

Helped by (measured better with it; the mesh does not wait on it):

• None in an open PR. A bot-side wedge escape, not in a PR, acts on the wedge listed under "Known, not changed".

At a glance

	master	this PR
nav areas	1463	1078
file size	1,792,918 bytes	1,030,012 bytes
nav ladders	3	3, linked up and down
areas no spawn reaches over any link	48 in 21 pockets	18 in 8
the same by each class's climb limit: Recon / Assault / Support	56 / 93 / 98	25 / 49 / 53
areas entered with no way back	3	0
objectives a team cannot reach	none	none (14 of 14, both teams)
fall deaths a match on neo master (A/B below)	2.46	1.28
bot stuck events a match on neo master	48.5	11.2
stuck episodes a match on neo master	8.8	2.6

What changed

• Areas the engine blocks at round start, where a brush or prop stands over an area's center, are cut around the
  obstacle or split, and standing room is carved around props where test bots stuck on them.
• The rooftop fall edges are marked. 35 of the 38 areas bots fell to their deaths from most often, along the
  rooftop edges, are PRECISE + CLIFF. The other three (9, 251, 4035) are large through-areas: marked, bots queued
  at one waypoint there (area 9: 37 stuck events a match) with no fewer falls, so they stay unmarked.
• Pockets with no way back are gone. Master has three areas bots can enter and not leave (3089, 3513, 4444); all
  three are removed.
• Walk links a player hull cannot cross are cut. The engine hull audit (a Recon hull swept along each link on 4 u
  lanes) found no clear lane on four links bots got stuck on: past an exterior wall trim, an ivy trim, and a large pipe
  prop. Two areas left with no way in are removed. Against the mesh before (10 pins): stuck events 101.7 -> 76.4 a match
  at reduced weapon damage.
  The audit of the final file cut more pairs both ways where no lane is clear either way: the nook 3654 to 408 near
  (1388, 475, -128) (a wall trim; 3654 keeps its link to area 73), a stair handrail at (-1212, 130), a tree at
  (-1875, 968), two vent supports at (112, 150) and (425, -38), and the large wire pipe at (625..760, 225..300), whose
  sloped end and a post split the areas on its top from the floor east of it. The ramp 4294 at (350..525, -600..-575)
  and the roof slab 12 north of it were linked across the 130 u trench between them, past a girder railing. The audit
  read 7 of 13 lanes clear there, but with the link back in, test bots logged 884 stuck events on the ramp in 20 of 20
  matches, so it is cut both ways.
• Strips inside props are removed: along the handrails of the z 170 deck, around a double post near (-300, -280), in a
  wall at (-2212, -117), a 25 u nook in a player clip at (788, 488), and four areas on a thin 45 degree plate 100 u above
  the floor at (1275, -440) that no spawn reached.
• Hand edits where test bots got stuck:
  • the clip near (1042, -1005): a player clip hangs 60 u above the floor at area 54's south edge, so a standing Support
    could not pass under it. The strip under it is a CROUCH area. With the ghost pinned at 6 spawns, 2 matches each:
    stuck events at the spot 26 -> 0, on the map 8.9 -> 6.0 a match, no new spot.
  • the stair corner near (-2265, -1900): bots skipped ahead diagonally under the stairs' low end, too low even
    crouched. The rail stretch of the area north of it is removed and its north piece is PRECISE, so paths run
    through the open area to the west.
  • the catwalk near (-760..-520, -435..-405, 0): an earlier pass merged master's 4268 into 4267, so 4267's drop to the
    floor 3909, 168 u below, ran under the catwalk 4441 / 4442 that crosses above 3909. A path dropping there aims at a
    point under the catwalk, and bots walked the catwalk back and forth toward it until the stuck monitor fired. 4267 is
    split at y -450 again as master has it: 4267 keeps the drop, 4268 links to the catwalk and the floor west of it only.
    On neo master, 50 matches a mesh: stuck events within 90 u, 2.68 a match (26 matches) before the split, 0.28 (10
    matches) after; master 0.06 (and 0.30 fell off the path).
• The rooftop strip near (-1520, 85, 221). Master covers it with area 3, 350 x 375 u. Here area 3 keeps its south
  part (y -150..-53), and the rest is two areas, the strip 4478 and 4471 east of it, linked both ways to each other, to
  3 and to the west stair. With the ghost pinned on the strip (ghost spawn 4), 6 of 6 matches picked it up on each
  mesh, median 44 s on master and 31 s on this one.
• Ghost spawn 1 near (-3317, 135, 43). The ghost rests on a case on a desk, and the engine picks a path's goal
  area with a line-of-sight test that the desk blocks for the floor area under it, so every path to the ghost ends in
  the 25 u strip along the west wall beside it. Master has that wall as eight strips; an earlier pass merged them into
  one 400 u strip, and a path from the south then crossed into it at its south end near (-3425, -80), 240 u from the
  ghost. The strip is split back at y 75, 125 and 200, so the area beside the ghost is 50 to 75 u long, as on master
  (3978 and 3981 keep master's ids and corners). With the ghost pinned there, 13 matches a mesh: pickup median 13.7 s
  on the mesh before this split and 10.7 s on this one (master 10.9 s, 6 matches), and bots ran to the strip's south
  end before the pickup in 13 of 13 matches before, 1 of 13 after.
• 11 rail-top slivers removed on the rooftop strip near (-1700, 0..225, 223) and (-1650..-1425, 262, 223), 10 x 25 u
  each: master linked them to area 3, and with 3 split they had no link in or out.
• Ladder 1 near (-1037, -376) is kept, with its two foot areas AVOID.
• Walk links restored: eight links that ran one way after earlier cuts get their missing direction back where the
  hull audit reads them clear both ways: three between level floor and five drops back from climbs. The climb at
  (-1012, 227) through a stair handrail is cut instead, and so is the 100 u climb from 12 to 701 near (150, 60), across a
  25 u gap, which no class plans.
• Marks: CROUCH 59 -> 30, NO_MERGE 118 -> 41, OBSTACLE_TOP 69 -> 15.
• Ids follow master's. 1062 of the 1078 areas keep a master id; 870 of them have master's corners. 16 areas are new.
• Re-analyzed in game after the edits, so the visibility lists bots use are current.

How it was tested

• Reachability of this file:
  • every spawn reaches every objective for both teams (14 of 14); no route is more than 10 % longer than master's
  • no dead end, no area a class can enter and not leave
  • by each class's climb limit, Recon does not reach 25 areas, Assault 49, Support 53 (by the bot planner's own limits, 27 / 27 / 59)
  • ghost spawn 10 at (-1047, 511) lies on a desk top, 52 u from the mesh: the floor area beside the desk now stops at
    y 499, where master's ran under the desk (35 u of headroom). With the ghost pinned there, 6 of 6 matches picked it up
    on each mesh, median 23 s on master and 22 s on this one. No area is added on the desk top: bots take the ghost from
    the floor beside it.
• Removed floor. 19 regions a spawn reached on master, 42,806 sq u, are gone: master's no-way-back pockets, ten
  strips at z 170 around (-1725..-1200, -200..300), 23,375 sq u, which lie along the deck's handrails, inside the rails'
  collision, floor cut back around the office chairs near (-3350..-2900, -1600..-1250), the stair corner's rail stretch,
  a rail top at z 214, the floor under the desk at ghost spawn 10, and two CROUCH strips near (-475..-400, 1175..1225,
  -256). Each of the two strips has a leg of a vent support (a 4 x 3 u post up to the duct at z -180) inside the box the
  engine tests over its center, so the engine dropped both from pathing for the whole round; they went with the other
  round-start blocked areas.
• Falls. The map has 8 trigger_hurt volumes: a pit, the roof, and six tanks. In the A/B below, 2.46 fall deaths a
  match on master's mesh and 1.28 on this one (paired ratio x0.64 [0.54, 0.76], lower in 35 of 50 pairs, higher in 7). From
  the 37 PRECISE areas (ids as master's): 1.06 and 0.18 a match.
• Engine hull audit of the final file (every level link swept with a player hull in game):
  BLOCKED_LINKS map=ntre_skyline_ctg level_pairs=1557 oneway=0 keep=2 cut=0 area=0 check=0 dynamic=0 entity=0 one_way_drops=194 sinks(any/recon/assault/support)=0/0/0/0. The two kept pairs read blocked and are walked: 617 / 871,
  the west stair up to the rooftop strip, whose lanes start in the slope (with the pair cut, no bot picked up the ghost
  pinned on the strip in 6 matches; master 6 of 6), and 31 / 633, the floor beside the desk at ghost spawn 10, the
  spawn's nearest area.
• Master against this mesh on neo master. Upstream neo master 092203f36 with no open PR, Release build plus the
  harness's logging. Byte-identical copies of the map, one carrying each .nav, matches interleaved: CTG, the ghost
  pinned at each of the 10 ghost spawns in turn, 5 matches each (50 a mesh), 30 bots, default weapon damage (1.485), 1
  round, every player at one rank tier rotated across the pins, 8 matches at a time. Stuck events are OnStuck plus
  FAIL_STUCK, without the spawn crush at round start and weapon scavenging. Paired ratio: geometric mean of
  (this mesh + 1) / (master + 1) over the 50 pairs, with a 95 % bootstrap interval.

a match, neo master	master	this mesh	paired ratio [95 %]
all events (stuck, fell off the path, no path, fall deaths) [95 % interval]	54.6 [45.1, 65.9]	15.7 [12.6, 19.8]	x0.30 [0.23, 0.37]
stuck events	48.5	11.2	x0.23 [0.17, 0.32]
stuck episodes	8.8	2.6	x0.37 [0.31, 0.43]
no path / fell off the path	2.06 / 1.56	1.82 / 1.34	x0.88 / x0.88
fall deaths	2.46	1.28	x0.64 [0.54, 0.76]
kills	24.8	25.4	x1.02
matches with the ghost picked up	48 of 50	50 of 50
matches at match end with 0 asserts	50 of 50	50 of 50

By rank tier, stuck events master / this mesh: Private 56.9 / 12.2, Corporal 37.0 / 13.5, Sergeant 50.2 / 7.1, Lieutenant
50.1 / 11.8. Ghost pickup median, master / this mesh: ghost spawn 1 (the desk) 12 / 12 s, ghost spawn 4 (the rooftop
strip) 46 / 26 s, ghost spawn 10 (the desk at (-1047, 511)) 22 / 23 s, all 5 of 5 on each mesh. In two more batches of
the same shape (master against this PR before the catwalk split, then master, that file and this one), all events went
49.9 -> 17.9 and 57.7 -> 18.3 a match; 250 more matches, all ended clean with 0 asserts.

Master's leading spots, events a match on master -> this mesh (neo master): the stair corner near (-2264, -1854) 12.4 -> 0;
ladder 1's foot near (-1009, -377) 7.6 -> 0.7; the office chairs near (-3074, -1405) 4.1 -> 0; the clip near (1000, -992)
2.0 -> 0; the double post near (-309, -289) 2.0 -> 0; the deck handrail near (-1731, 0, 170) 1.5 -> 0; the stair
handrail near (-968, 311) 0.4 -> 0.1. The helicopter prop on the rooftop strip: 0.37 on master, 0.18 on this PR's mesh
over the three batches. This mesh's leading spots are master's as well, listed under "Known, not changed": the slope south of
area 73 (3.3 a match here, 1.0 on master in this batch) and the pipe prop at area 675 (2.8, master 2.3).

Earlier measurement on the harness's research build (neo master plus the open bot PRs of the time and further research bot changes), before the catwalk split

a match	master	this mesh
all events (stuck, fell off the path, no path, fall deaths) [95 % interval]	32.9 [25.5, 40.8]	4.3 [3.1, 5.6]
stuck events	28.9	2.4
stuck episodes	5.2	0.8
no path / fell off the path	0.37 / 2.40	0.25 / 0.77
fall deaths	1.20	0.87
kills	25.5	25.2
matches with the ghost picked up	56 of 60	57 of 60
matches at match end with 0 asserts	60 of 60	60 of 60

That table is master against the mesh before the ghost spawn 1 split and the sliver removal, 10 pins x 6. Those two
edits were then measured against that mesh the same way, 10 pins x 5: all events 4.2 -> 4.3, stuck events 2.6 -> 2.5,
fall deaths 0.84 -> 0.88, ghost picked up 49 -> 48 of 50, 50 of 50 clean each.

Known, not changed

• The pipe prop at area 675 near (575, 200, -130): bots heading south wedge on the pipe at (530, 220, -113), west of
  the area, against the world, in about 2 of 5 matches on either mesh. The areas and links there are the same on master
  and here (this mesh only marks 675 PRECISE). On neo master over three batches: 2.8 events a match on master's mesh
  (150 matches), 3.6 on this PR's mesh before and after the catwalk split (200), the same move each time; 0.02 on the
  research build, whose bot code the spot depends on.
• The slope south of area 73 near (1450, 540, -130): a bot entering 3667 from 73 circles the path point at
  (1450, 540, -128) until the stuck monitor fires, on either mesh; the areas 73, 3656 and 3667 and their links are master's.
  Over three batches on neo master: 2.7 events a match on master's mesh, 2.8 on this one, in 44 and 69 of 150 matches.
  Bots cross from 73 into the slope 1.6 times a match on master's mesh and 2.6 on this one, so events per crossing are
  lower here (1.1 against 1.7); the rate follows traffic.
• A wedge off the roof slab's south edge near (110, -528, -21): a bot crossing from area 12 to 400 can drift off the
  edge and wedge in the air between two round props for the rest of the round. 5 times in about 3,100 matches.
  In a targeted test, marking the edge PRECISE kept bots on it (position samples off the edge 6 -> 0, 24 matches an
  arm), but they then stood and fought on the marked strip twice as long; the mark is left out of this mesh and the trap
  left to bot code.
  On neo master: no event and no fall death within 80 u on either mesh in 350 matches (three batches).
• Four one-way drops where master has the climb back: 987 -> 3919, 1102 -> 82 and 1102 -> 3915 near
  (338..425, 325..425, 48), 48 u, and 4464 -> 4465 near (-1670, 790, 35), 39 u. Each has a planned way back up nearby.
  On neo master, 350 matches: one bot stepped down 1102 -> 3915 and back up; 4 events within 110 u in 200 matches on
  this PR's mesh, none on master's.
• 7 areas near (-38, -359, 110) reached only over a climb of 109 u, the same on master.
• The stair corner at area 253 near (-2265, -1900): three edits were tried; the one above is kept, the gap between
  the rail and the stair's post has no nav area. On neo master: 12.4 events a match within 100 u on master, none on
  this mesh.