collision-hull-proxy

Collision Hull Proxy

A runnable example that builds a fire hydrant street prop — lathed barrel, bonnet dome, three outlet caps, operating nut — plus the collision a game prop pipeline ingests: a compound of four convex hull pieces, one per part group, each built with bmesh.ops.convex_hull from a coarse collision cage, following mesh-editing-and-bmesh.

Pipeline arc neighbor: LODs in lod-decimate-chain, tangent space in triangulate-tangents, export in gltf-export-roundtrip — collision bounds are the piece an engine's physics ingest checks first, and the piece whose failure rejects the whole prop.

Scope: this witnesses the bpy-level contract a prop pipeline relies on (watertight convex pieces that enclose the render mesh within budget). It is not an engine exporter and does not produce a shippable FiveM/engine asset — it proves the geometry properties such an asset must have.

What it witnesses: engines (GTA/FiveM-style prop workflows included) ingest collision as a compound of convex pieces, each piece under a per-piece face limit (255 is the common cap). The example builds that structure and derives every property from closed forms, per piece:

• Containment. Every render-mesh vertex lies on the inner side of every face plane of its piece's hull — max excursion 4.4e-08 (tol 2e-4). This holds by construction, not luck: the cage is a coarse lathe whose rings are inflated by sec(π/n), so each n-gon cage ring circumscribes its render ring exactly. Proud details (a rim, a nut) must be paid for in cage rows; concave details (the grooves) are free — the hull bridges over them. That trade-off is the collision-authoring lesson.
• Convexity. The same plane test restricted to the hull's own vertices (4.8e-08) — what a convex solver assumes when it uploads the piece.
• Watertight + manifold. Every hull edge borders exactly two faces, and V − E + F = 2: a convex hull is a topological sphere.
• Outward winding. Positive signed volume by the divergence theorem (body piece 1.022381) — inverted collision is a physics no-op.
• Budget. Every piece ≤ 255 faces: body 70, pumper 60, side caps 60+60; the compound totals 250. The naive approach fails this — a hull of the dense render mesh measures 380 faces, over budget, which is why real pipelines hull a cage, never the render mesh.

What each check catches on failure: an artist editing the render mesh after the hull was generated (probe: pumper lug stretched 2×, exit 3, measured escape 0.730000); inverted hull winding (probe: polygons flipped, exit 5, signed volume −1.022381, the exact negation); a non-watertight piece (probe: one face deleted, exit 4, three border edges).

Version witness: output is byte-identical on Blender 4.5.11 LTS and 5.1.2 — same piece counts, same volumes, same excursions. bmesh.ops.convex_hull and foreach_get are stable across the pair.

The render shows the prop inside its four faceted translucent hull pieces: if any piece failed to enclose its geometry, painted metal would poke through the shell in frame.

Run

# Cheap correctness check (no render) — the CI check:
blender --background --python collision_hull_proxy.py --

# Also render a still (EEVEE on a GPU host; use --engine cycles on GPU-less hosts):
blender --background --python collision_hull_proxy.py -- --output hydrant.png
blender --background --python collision_hull_proxy.py -- --output hydrant.png --engine cycles

It exits non-zero on failure (render geometry escaping a hull, inverted winding, a non-watertight or non-convex piece, Euler drift, or a piece over the 255-face budget). The blender-smoke workflow runs the check on Blender 4.5 LTS and 5.1.