For researchers: What physical interaction scenarios are underrepresented in world model training data?

Hello, I am a 3D artist with a Unity (game engine) background, thinking about producing synthetic data targeted at world models, physical AI training, or just spatial intelligence. Before I build anything, I want to hear from people actually training or evaluating these models about where the real gaps are. My current read of the landscape is that NVIDIA and the major lab pipelines already have the high-volume cases well covered, meaning street scenes, warehouses, generic environments, humanoid locomotion. If you have a minute: (1) - What physical scenarios do you wish you had more of in training or eval sets? (2) - Where are current world models visibly failing in your experience? (3) - Are there benchmarks you wish existed but don’t? I am considering doing a relatively small but focused dataset. Something that could be tested and hopefully brings someone value. I am also looking to learn, so feel free to drop in any knowledge I could benefit from.