Somewhere in a dungeon you've played a hundred times, you clip through a torch bracket mounted on the wall. Nothing breaks. You don't fall through the floor. The game just nudges you sideways and carries on, and your brain files the whole thing under "fine" before you've even processed it.
That tiny moment is a collision system doing its job, which mostly means deciding what not to simulate.
The invisible geometry underneath everything
Every object you can see has two versions of itself living in the engine simultaneously. The visual mesh, which might carry 50,000 polygons on a detailed enemy character, and the collision mesh, a handful of crude geometric primitives stacked together like a rough sculpture built from cardboard boxes. A human character's collision body is typically a capsule: a cylinder capped with hemispheres at each end, because capsules slide along stairs and ledges without snagging. The art team can spend weeks perfecting the visual silhouette. The physics team wraps it in a shape a four-year-old could describe.
This isn't laziness.
It's load arithmetic. A physics engine checking whether two objects overlap runs that test every frame, sixty or more times per second, against every collidable object in range. Checking two capsules for overlap is a handful of floating-point operations. Checking two 50,000-polygon meshes is millions, and at scale, with dozens of characters, projectiles, and debris objects all moving at once, the difference between a fluid 60fps and a slideshow is often just how aggressively the collision geometry has been simplified.
The rule most engines follow is roughly this: if a player can't meaningfully interact with a surface detail, it doesn't need a collision boundary. That torch bracket? Decorative. The engine won't miss it. The chest a player opens to grab loot gets a precise box collider, because players click on it and expect a specific response.
Why some shortcuts feel fine and others feel broken
Consider a stone archway, roughly two metres wide at the base, with carved decorative edging that juts out about eight centimetres on each side. A physics programmer has two options: build a collision mesh that follows those carved edges exactly, or use a single rectangular box covering the full opening and ignoring the decorative detail. If the player is running through at speed, they'll never notice. The eight-centimetre gap between the real visual surface and the invisible collision box is smaller than the margin of error in how most players steer.
The term for that margin is the collision tolerance, and most third-person action games use tolerances between two and fifteen centimetres depending on gameplay context. Tight platformers run tighter tolerances because players are making millimetre judgements about ledge edges. Open-world games run looser ones because the sheer volume of environmental geometry would otherwise tank performance.
What breaks the illusion isn't inaccuracy. It's inconsistency. Two players, call them Priya and Tom, both buy the same action RPG. Priya runs along a fortress wall, clips slightly through a battlement, and keeps moving. She never registers it. Tom does the same thing three times near a narrow gate, bounces back each time, and decides the game is janky. The underlying collision data is identical. The difference is that the gate is a meaningful chokepoint and the battlement is scenery. When simplified geometry contradicts something the player is actively trying to do, the shortcut becomes visible.
Engines handle this with collision layers and priority flags. Objects tagged as gameplay-critical, doors, platforms, weapon-impact surfaces, get higher-fidelity meshes or even exact convex hull approximations computed automatically by the engine's physics toolchain. Unreal Engine's built-in convex decomposition tool breaks a complex concave shape into a set of simpler convex pieces that together approximate the original within a configurable error threshold. Unity's PhysX integration does something similar. The programmer sets the acceptable error; the tool finds the simplest geometry that stays within it.
The part that gets underappreciated: "simplest geometry that works" is a judgment call that shifts with camera angle and player speed. Speed hides a lot. Slowness reveals everything, the way a slow zoom on a movie set will eventually find the edge of the backdrop.
Ever found a spot where you can clip through a wall if you approach at exactly the right angle? You've just found the seam between a gameplay-critical surface and a decorative one. The engine made a call. It was almost certainly the right call for the other 99.9% of players who will never stand in that exact spot moving at that exact speed.
The physics team knew it was there. They shipped it regardless. That's not a bug, it's a budget.