The researchers hope to create a 'reliable and controllable video game generator,' and possibly replace traditional game development entirely with video generation models.
Using a public dataset of Super Mario Bros. gameplay, the MarioVGG model was trained on over 737,000 frames to learn physics and gameplay dynamics through limited inputs.
Despite current limitations, the MarioVGG model demonstrates that AI can infer impressive gameplay dynamics by learning from just a limited set of video and input data.
The project reflects an innovative step in AI-driven game generation, although the output still exhibits glitches and lacks real-time functionality.
Collection
[
|
...
]