Continuous Collision Detection (CCD) vs Discrete Collision at Hyper Velocity: Edge Cases, Failure Modes, and Root-Cause Debugging

[Alex Young]

I’ve been hitting a wall with tunneling on fast projectiles in our custom physics pipeline. When velocities exceed a threshold, discrete swept-SAP tests start missing narrow targets entirely, and I’m getting ghost collisions right through thin geometry. My current approach uses SAT on the swept capsule

"Dota 2 Immortal Bracket | Position 1 Carry • 7.5k MMR"

[Isaac Smith]

Great topic — this is exactly the kind of physics edge case that ruins matches in fast-paced shooters and fighting games. The core issue with discrete collision is tunneling: at high velocities, an object can pass entirely through a collider between two simulation frames because the engine only samples positions at frame boundaries. If your bullet moves 40 units per frame and the wall is 20 units thick, the sweep never lands on a frame where the bullet overlaps the wall, so no collision is registered. CCD solves this by sweeping the volume along the trajectory and finding the earliest point of intersection, which is why most competitive titles enable CCD (or continuous sweep-based raycasts) for fast projectiles specifically. The catch is that CCD is more expensive and can introduce its own artifacts, so it’s usually gated to high-velocity objects rather than applied globally.

The nastiest failure modes I’ve hit are CCD false-positives on thin geometry, “slop” jitter where the solver repeatedly resolves a nearly-grazing sweep, and the classic case where CCD is on but the collision normal is computed wrong, sending the projectile sideways instead of stopping it. For root-cause debugging, I always check three things: the per-frame displacement versus collider thickness ratio, whether the object’s collision layer/mask is actually being swept, and if the broadphase (usually an AABB/SAH) is culling the sweep because its swept bounds got miscalculated. A quick heuristic is to log the object’s velocity magnitude at the frame of the miss — if it’s above the threshold where tunneling becomes likely, you know discrete was the culprit. Enabling CCD

"Overwatch 2 Top 500 Hitscan | 1600 DPI @ 2.5 Sens • EU Region"

[Author]

Posts