Six open video world models drift toward human eye height, new EyeHeight benchmark shows
Starting at 0.15 m, only 0-20% of clips from the six models end within 1.3 times of the starting height after a 180-degree turn-and-walk rollout.
What happened
Some AI models act as simulators for robots. They take a starting image and a list of moves, then generate what the robot would see. Researchers found that six open models slowly shift the camera toward human eye level, even when the robot's camera sits near the floor.
Why it matters to you
A simulator that changes the camera height gives wrong training and test views to robots, so their planned paths can be off.
The takeaway
AI world simulators can forget how high the camera is, which makes them unreliable for low or high robots.
What it means for you
If you use world models to simulate low or overhead cameras, check height retention with a benchmark like EyeHeight before trusting the rollouts.
Robotics teams that train or test agents in generated simulations may get views that do not match their real camera mounts.