This decision happened before I joined Epic, but I believe a principle reason was that we wish to be able to modify the physics engine so that it more fits our needs. Some of the things we want to build require deep integration into the physics engine which would be difficult or impossible when using an external physics engine. E.g., destruction, networked physics, etc.
Chaos is actually based on PBD. This comes with problems of its own and is not quite the nirvana one might think, but it works and we’re running with it. One restriction is that we must compete with PhysX in terms of performance on existing scenes, which means we have had to compromise some benefits of PBD for this. E.g., we have linearized the constraint error calculations. One day we will provide options to select between linear and non-linear, but not yet.
Our ultimate goal is to support cross-simulation of rigids, soft bodies, cloth and fluids, but that’s a fair way off. In the meantime we have research projects on various ways to improve simulations, but none are ready for discussion at this time.