The idea

Models apply forces, query object states, advance time and change the scene in MuJoCo before answering, and the physics engine checks the result. Much of what a task needs isn't visible up front. It can only be found by experimenting, watching what happens and adapting.

The finding

Models reached up to 80% on qualitative single-concept tasks, but most fell below 30% on the hardest. The failures were mostly about procedure: answering too early, exploring inefficiently, and not grounding in what the simulator actually showed.