Rehearsing physical tasks before trying them for real
Robotics and embodied AI are where world models arguably matter most. Real-world robot training is expensive, slow, and can involve real risk: a robot arm that fails while learning to grasp a fragile object can damage it, and a legged robot that falls while learning to walk can damage itself. A robot equipped with a decent world model can rehearse a task, like a grasp, a step, or a manipulation sequence, inside its learned simulation first, filtering out plans likely to fail before ever attempting them on hardware.
This connects directly to material on reinforcement learning for robot control, but the emphasis is different. Rather than learning a control policy purely through repeated physical trials, a world model gives the robot an internal sandbox. That sandbox is cheaper to run, safer to fail in, and can be explored far more times than real hardware ever could be in the same span of time.