I would guess that pretty much all vehicle models use either a single rigid body or multiple rigid bodies. They will all need to know the centre of mass. Even military simulations use vehicles based on rigid bodies. Additionally, PhysX isn’t too far from CarSimEd, which was also based on rigid body mechanics. Anything that computes tire or suspension forces will need to apply those forces to something that is almost certainly a rigid body.
The core of the model is that there is a rigid body and there are forces arising from tire and suspension forces. The rigid body’s transform helps compute the forces and the forces tell the rigid body how to move. I’m guessing that pretty much all vehicle models work on that principle and only differ to the degree of complexity of the model. Some models have more sophisticated tire models, some model the gears and clutch more accurately, some model the suspension system with multiple bodies coupled with constraints. PhysX does all of these in a fairly simple way but it shares the core model of a rigid body moving under force.
The problem with all simulation is that the model used is always an approximation. I wouldn’t even say that the centre of mass estimation is the biggest approximation in a vehicle simulation. Tire forces, for example, are always going to be approximate because it is hard to mesure the regression data for a specific tire on a specific car on a specific road in specific temperature conditions. Even if we had a great model it would be very difficult to drive that model with accurate data. The suspension model is also prone to error. PhysX uses a simple spring but it would be easy to extend that to a jointed system that more closely mimics reality. If we did that we would be faced with all sorts of technical problems like how to accurately model a non-linear spring or one that has bi-directional damping. Even measuring the stiffness graph of a non-linear spring is a significant challenge. How would we even measure that data so that it could be plugged into a vehicle simulator? Finally, there is the timestep of the simulation. As the timestep reduces the simulation ought to to converge but that requires very, very small timesteps that are computationally expensive. Making a good timestep choice that balances computational expense with accuracy is fraught with difficulty. This is all hard.
It is always difficult to map the real world to physics modelling parameters. That is true of all vehicle models as well as PhysX. I would say that PhysX is a fairly simple model; more sophisticated packages will ask for more and more parameters that might not be readily available to the user. The best we can do is to improve our estimates by calibrating the simulated motion against real data.