Physics can get squirrelly in multiplayer, where two different players will think they're pushing the same blocks in different directions. Try piling 200 cubes and spheres of different sizes into a room and have six players walking around and see whether it gets glitchy.

You also have the challenge of things where players need to collaborate. If you're on a network, having two players both attached to one rigid body will not work out well, because of the latency. In fact, even attaching a single player to a rigid body has problems, but those are solvable by trusting the client for that one body. (Until that body collides with a second player.)

The problem is much more fundamental than that. You have a dependency graph of who affects what joints/bodies in the physics simulation, and the graph has two different latencies/step offsets on the two different client machines. Unless you delay all client simulation until all inputs for all other clients are known for a given step (which generates a lot of control latency,) this will always generate pretty severe physics differences.

Note that even just “colliding” with another rigid body will create a joint between the colliders (the “contact joint”)