NarrowPhase

Narrowphase refers to the stage in collision detection and physics simulation where precise tests are performed on pairs of objects that the broad phase has flagged as potentially intersecting. The goal is to determine whether a real collision occurs and to produce detailed contact information that can be used by the physics solver.

In contrast to the broad phase, which uses simple bounding volumes to quickly prune unlikely pairs, the narrow phase applies accurate geometric tests. For convex shapes, common methods include the Gilbert-Johnson-Keerthi (GJK) algorithm to detect intersection and the Expanding Polytope Algorithm (EPA) to compute penetration depth and the contact normal. The Separating Axis Theorem (SAT) and various feature-based approaches are also used. For non-convex shapes, the problem is typically reduced to testing convex components through decomposition, Minkowski sum techniques, or other approximations.

The primary outputs of the narrow phase are the contact manifold and associated data. This typically includes one or more contact points, each with a contact normal, penetration depth, and sometimes relative velocity or time of impact information for continuous collision detection. This information is fed into the constraint solver to compute impulses, resolve interpenetration, and apply realistic response forces.

Workflow and relationship to the broad phase: the broad phase rapidly culls non-overlapping pairs using spatial partitioning structures such as bounding volume hierarchies, grids, or sweep-and-prune; the narrow phase then performs the exact tests on the remaining candidates.

Performance considerations are central in narrowphase design. Techniques include caching and warm-starting contact data across frames, reusing previously found contact points, parallel processing, and selective detail based on object motion. Handling complex or highly concave shapes and high-velocity motion often requires decomposition, approximation, or alternative algorithms.