World Emulation via DNN

Ollin Boer Bohan has developed a neural network-based "neural world" that allows users to explore a forest trail through their web browser. This innovative approach utilizes a neural network to generate images based on previous frames and user controls, without traditional game development elements like level geometry or scripted animations. The project builds on Bohan's earlier work, where he created a 2D game world by training a neural network on gameplay videos. The new neural world was created by recording approximately 15 minutes of video while walking through a forest, which was then processed into input-output pairs for training. Despite initial challenges in achieving a coherent output, Bohan made several upgrades to the training process, including enhancing control inputs, increasing memory capacity, and refining the network architecture. The final model, which consists of around 5 million parameters, was trained on a dataset of 22,814 frames. Bohan argues that while traditional game worlds are crafted like paintings, neural worlds are akin to photographs, capturing real-life details directly from recorded data. He envisions a future where neural worlds can achieve high fidelity and become a distinct creative medium, similar to how photography evolved. Bohan encourages programmers interested in world modeling to explore existing frameworks and expresses a desire to continue improving this technology.

- Bohan created a neural world that generates images based on recorded video and user controls.

- The project builds on previous work with 2D game emulation using neural networks.

- Significant upgrades to the training process improved the quality of the generated world.

- Bohan compares neural worlds to photographs, emphasizing their basis in real-world recordings.

- He envisions future advancements that could make neural worlds a unique creative medium.