Physics Collider Optimization in Complex Game Scenes

id: 240 slug: physics-collider-optimization-in-complex-game-scenes title_en: "Physics Collider Optimization in Complex Game Scenes" tags: [vr-ar]

Physics Collider Optimization in Complex Game Scenes

Scene with 800 GameObject, each with MeshCollider by real geometry — PhysX spends 8–12 ms on FixedUpdate only on Broadphase and Narrowphase checks. This half frame-budget on physics alone. In VR where frame budget 11 ms (90 FPS), immediate drop. Collider optimization — one of those tasks where correct architecture from start saves weeks of work later.

How PhysX Spends Time and Where Losses

Physics in Unity passes two stages: Broadphase (fast filter: which objects even close to each other?) and Narrowphase (precise intersection check for filtered pairs).

Broadphase uses AABB-tree (Axis-Aligned Bounding Box). If objects constantly move — tree constantly rebuilds. 200 Rigidbody with isKinematic = false, moving each frame, give constant AABB-tree rebuild. Solution: objects not moving now, switch to isKinematic = true via script — they fall out dynamic tree.

Narrowphase — where MeshCollider kills performance. Intersection check of two arbitrary meshes — O(n×m) by triangles. Character collider against building collider with 50,000 triangles — thousands of operations per Narrowphase. PhysX doesn't support non-convex MeshCollider against MeshCollider in dynamic objects at all — only against Static.

Practical Replacements and Simplifications

Rule one: replace MeshCollider where imperceptible to player. Chair with MeshCollider by real geometry — 200 triangles. Same chair with 2 BoxCollider (legs + seat) — 2 primitives. Collision accuracy 95% of original, cost — 50x less.

For organic shapes (rocks, barrels, cars) set of 3–6 ConvexMeshCollider (convex simplified meshes) works orders of magnitude faster than non-convex and visually indistinguishable. In Unity: MeshCollider.convex = true + simplified mesh via LOD or manual collision proxy creation in Blender.

Compound Colliders — set of primitives as child objects of one Rigidbody. Only way to get "non-convex" dynamic physics without non-convex MeshCollider: several convex pieces = complex shape. For weapons in VR, detailed objects, robots — standard approach.

Collision Matrix (Layer Collision Matrix). In Physics Settings → Layer Collision Matrix disable checks between layers never interacting: Environment vs Environment, UI vs Physics Objects, VFX vs any layer. Each disabled pair — fewer Broadphase pairs. Proper matrix setup reduces active pairs by 30–50%.

Sleep Threshold. Rigidbody sleeps when velocity drops below Physics.sleepThreshold. Sleeping Rigidbody doesn't participate in physics — near-zero cost. Default value (0.005) too low for complex scenes: objects stay active on microvibrations. For VR-scenes without fluid/cloth simulation set 0.1–0.2.

Diagnostic Tools

Physics Profiler in Unity Profiler — shows Broadphase, Narrowphase time, separate SimulateAndCollide, UpdateTriggers. Here visible where losses.

Physics Debugger (Window → Analysis → Physics Debugger) — visualizes colliders in Scene View with color-coding by type (Static, Dynamic, Kinematic, Trigger). Quick identification of unexpectedly heavy colliders.

PhysicsViewer custom Editor Tool — script outputting list of all MeshColliders in scene, sorted by triangle count. Written in 30 minutes, saves hours of searching.

From practice: in VR-trainer for industrial equipment, plant shop scene contained 1200 objects, most with MeshCollider by imported CAD-models (15,000–80,000 triangles each). Physics CPU time — 14 ms. After replacing all static MeshColliders with Convex proxy (auto via Editor script + VHACD decomposition) and disabling unneeded Layer Matrix pairs — Physics CPU time dropped to 2.8 ms.

Work Stages

Profiling. Physics Profiler snapshot, bottleneck identification.

Scene audit. All colliders inventory, classification by type and cost.

Optimization. MeshCollider replacement with primitives/Compound, Layer Matrix setup, Sleep Threshold.

Automation tools. Editor scripts for batch-replacement and validation.

Re-profiling. Comparison with baseline, changes documentation.

Cost calculated after profiling and change scope assessment.