Aftermath

PointCloudShape vs ConvexHullShape

In this example, we used the PointCloudShape to simulate a rigid body with a convex hull. As the name suggests, this shape only requires a set of points — they don’t need to lie exactly on a convex surface. By design, Jitter treats the convex hull of these points as the actual shape for collision detection.

Although the algorithm used here is brute-force, it can be extremely fast: the data is stored in a linear memory layout, and SIMD instructions are used to accelerate the support function. Because of this, PointCloudShape is usually the best choice for quickly and efficiently adding simple convex geometry to your simulation.

The ConvexHullShape, on the other hand, is intended for more complex and detailed convex models. Unlike PointCloudShape, this shape requires a precomputed convex hull provided as a list of triangles (List<JTriangle>). Jitter does not generate this for you — you'll need to use third-party tools like Blender or dedicated convex hull libraries. The input mesh must be strictly convex for collision detection to work correctly.

Internally, ConvexHullShape uses a hill-climbing algorithm to compute the support function. While this approach is algorithmically more efficient than brute-force, the performance benefits only become noticeable with larger shapes. As a general rule of thumb, ConvexHullShape starts to outperform PointCloudShape at around 300 vertices or more.