[sunzenshen]

Description

Look along the ledge for a drop-down's clear column

Background

This came out of the navmesh work on ntre_sentinel_jgr. Bots whose path dropped through a roof hatch walked past the hatch onto solid floor, pressed into what lay beyond until the stuck monitor fired, and then got the same path again: 44.9 stuck events a match at the hatch. The same happens wherever a drop's opening is off to one side of where a path crosses the ledge, such as a gap in a railing, so it is fixed in the path code rather than in one mesh.

Path::PostProcess turns a walk link onto a lower area into a DROP_DOWN. It does this by pushing the crossing point forward (10 u steps, up to two hull widths) until a hull traced straight down is clear.

It only looks straight ahead of where the path crosses the edge. When the opening below is off to one side of that line (a hatch in a floor, a gap in a railing), no column is clear. The loop still sets the drop point at the last, unchecked position: two hull widths out, on solid floor. The bot walks there, stands on the floor it was meant to fall through, and presses into whatever is beyond until the stuck monitor fires; the next path puts it back.

When the straight search finds nothing, the same search is now tried at offsets along the portal: either side of the crossing, nearest first, staying inside the portal. The path crosses at the first clear column. Paths whose straight search succeeds are unchanged. One file, NextBotPath.cpp, +42 inside #ifdef NEO.

Dependencies and status

Needs (code PRs this one depends on):

• None.

Helped by:

• None.

Pending before review:

• A match comparison on upstream master with and without this PR alone at the roof hatch, on master's ntre_sentinel_jgr mesh (so far measured as a research patch).
• A cold review read (none recorded yet).

Navmesh PRs that rely on this

None of the open navmesh PRs needs this: its case, the roof hatch, is on ntre_sentinel_jgr, whose navmesh PR NeotokyoRebuild/neoAssets#165 fixes the hatch in the mesh itself. It is still worth merging on its own: the defect can happen at any drop on any map, and the 10-map check below shows no harm.

Evidence

• The hatch case: ntre_sentinel_jgr roof hatch, on a mesh whose drop link spans the whole edge. Stuck events at the hatch went from 44.9 to 2.1 a match, and bot stuck events overall x0.55 [0.42, 0.71] over 36 paired matches.
• Do-no-harm, 10 maps on current meshes, 258 paired matches:
  • stuck events x0.87 [0.78, 0.98], mostly unrelated heavy tails
  • kills x1.01 [0.99, 1.04]
  • stuck events at the changed drops about 0 in both arms

Risk

Low. Only paths whose straight-ahead drop search already failed are affected. If no column along the portal is clear either, behavior is as before.

Toolchain

• Linux GCC Distro Native (Linux Mint 22.3, GCC 13.3.0)

Linked Issues

• related ntre_sentinel_jgr: navmesh rework neoAssets#165

[sunzenshen]

Can be solved with a navmesh edit.