Simulated robot skill tests count each failure against the robot's control program. But many failures are not the robot's fault, they come from simulation glitches that eject objects after correct placement.
To find such glitches, the researchers built a tool that flags any object accelerating faster than possible. The tool traced the glitch to a single incorrect constant in the PhysX physics engine, which misreports collisions between close objects. Correcting that constant fixes the faulty collisions.
With the fix, all simulator-caused failures disappeared, and the robot's success rate on container tasks rose from 89% to 98%. This shows that standard simulation scores can be misleadingly low.