DOOM Recreated Entirely By AI - Google GameNGen - The Future of GameDev or Just More AI Hype?

Gamefromscratch
29 Aug 202412:27

TLDRGoogle's GameNGen project is recreating the 1993 game Doom using AI, specifically a neural network and a modified version of stable diffusion for rendering. The AI learns gameplay aspects by analyzing 900 million frames of Doom, yet lacks a true understanding of game logic like hit scanning or inventory. While impressive, it's not a replacement for game engines and has limitations such as lack of permanence and memory constraints. The project hints at a future where games might be developed and edited through AI, potentially reducing costs and increasing accessibility.

Takeaways

  • 😲 Google's research project, GamenGen, recreates the 1993 game Doom using AI.
  • 🧠 The AI utilizes a neural network and a modified version of stable diffusion for rendering.
  • 🎮 AI learns gameplay aspects such as opening doors and needing keys, but lacks concept of hit scanning or inventory.
  • 🚫 AI can only remember the last few seconds, so objects like barrels can reappear after being destroyed.
  • 👾 The AI has been trained on 900 million frames of Doom gameplay at a low resolution.
  • 📈 AI extrapolates game frames based on recent actions, simulating the process of creating Doom levels.
  • 💻 GamenGen runs at over 20 frames per second, showcasing real-time interaction.
  • 🚧 The project has limitations, including lack of permanence and glitches in rendering.
  • 🔮 Future work may include testing the AI on other games or interactive software.
  • 🌐 While impressive, GamenGen is not a replacement for traditional game engines or developers in the near future.
  • 🤖 The technology could potentially lead to new paradigms in game development, making it more accessible.

Q & A

  • What is the name of the AI project that recreates Doom?

    -The AI project that recreates Doom is called GamenGen.

  • Which companies are involved in the GamenGen project?

    -The GamenGen project is primarily a research project from Google, with contributions from Tel Aviv.

  • What technology is used to recreate the Doom game in GamenGen?

    -GamenGen uses a neural network and a modified version of stable diffusion for the rendering side of things.

  • How does the AI in GamenGen learn to recreate Doom?

    -The AI learns by playing thousands of games of Doom and encoding 900 million frames of graphical data and logic into its neural network.

  • What are some gameplay aspects that the AI has learned?

    -The AI has learned aspects such as doors opening when entered and the necessity of a key at certain points in the game.

  • What is a limitation of the AI's memory in GamenGen?

    -The AI can only remember the last few seconds of gameplay, which means objects like barrels can reappear if the player leaves and returns to the area.

  • How does the AI render the frames in GamenGen?

    -The AI uses the modified stable diffusion model to render the frames based on the player's recent actions.

  • What is the frame rate of the AI-generated Doom game?

    -The AI-generated Doom game runs at over 20 frames per second.

  • What is the significance of the 900 million frames of training data?

    -The 900 million frames of training data at a low resolution are used to train the AI to understand and recreate the gameplay of Doom.

  • What are some of the limitations mentioned in the GamenGen project?

    -Some limitations include the lack of permanence in the game state, differences in behavior between the AI agent and human players, and the inability to explore all game locations and interactions.

  • What is the potential future application of technology like GamenGen in game development?

    -While GamenGen is not a replacement for game engines, it could potentially be used to create modifications or novel behaviors for existing games in the future.

Outlines

00:00

🎮 Introduction to Game Engine Project

Mike introduces a new research project named 'Game n Gen,' a game engine developed by Google and Tel Aviv. The project focuses on recreating the 1993 game Doom using a neural network and a modified version of stable diffusion for rendering. The engine learns gameplay aspects such as opening doors and needing keys but lacks concepts like hit scanning, inventory, or game state. It can only remember the last few seconds of gameplay. The engine uses a bot to play thousands of games of Doom, capturing 900 million frames of graphical data and logic to train its neural network. The rendering is done in real-time at over 20 frames per second, but the engine's lack of permanence and understanding of game mechanics like inventory and health are highlighted. The project is not intended to replace traditional game engines but showcases the potential of AI in game development.

05:01

🔍 Deep Dive into Game Engine's Training and Limitations

The video delves into the training process of the Game n Gen engine, which involved agents collecting 900 million frames of data at a low resolution. The engine extrapolates the next frame based on the previous 30 frames of gameplay. It uses a modified version of the stable diffusion model for rendering and additional logic to handle specific in-game scenarios. The limitations of the engine are discussed, including its inability to maintain object persistence and the difference in behavior between the AI agent and human players. The engine does not explore all game locations, leading to areas with non-existent or erratic behavior. The creators plan to test the engine on other games and interactive software in the future, aiming to improve its capabilities and reduce the cost and accessibility of game development.

10:02

🚀 Future Implications and Critiques of Game Engine

The video concludes with a discussion on the potential future of the Game n Gen project and its implications for the gaming industry. The creators are optimistic about the paradigm shift from traditional game engines to AI-driven systems, suggesting that it could reduce development costs and make game creation more accessible. They envision a future where games could be developed and edited through textual descriptions or example images. However, the video also presents a balanced critique, noting that the engine's current capabilities are limited and that it does not truly understand the game mechanics it mimics. The technology is compared to 'cosplaying' a game rather than being a functional game engine. The video suggests that while the project is impressive, it is not yet ready to replace traditional game development methods.

Mindmap

Keywords

💡AI

Artificial Intelligence (AI) refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the context of the video, AI is used to recreate the classic game Doom, showcasing its ability to learn and mimic gameplay aspects. The AI system is trained on millions of frames of Doom gameplay, allowing it to render and simulate the game environment in real-time.

💡Game Engine

A game engine is a software framework designed for the creation and development of video games. It provides the necessary tools and functionalities for game developers to build and run their games. In the video, 'Game n Gen' is a research project that aims to use AI to recreate the functionality of a game engine, specifically for the game Doom, by learning from game data rather than traditional programming.

💡Doom

Doom is a classic first-person shooter video game released in 1993. It is known for its fast-paced gameplay and has been influential in the development of the genre. The video discusses a project where AI is used to recreate the game Doom, demonstrating the AI's ability to learn and replicate gameplay mechanics, level design, and other aspects of the original game.

💡Neural Network

A neural network is a series of algorithms modeled loosely after the human brain that are designed to recognize patterns. It is a crucial component of AI and machine learning. In the video, a neural network is trained on millions of frames of Doom gameplay to enable the AI to recreate the game. The network learns from the data and makes predictions about what should happen next in the game based on the input it receives.

💡Stable Diffusion

Stable Diffusion is a model used in AI for generating images from text descriptions. In the context of the video, a modified version of Stable Diffusion is used for the rendering side of the AI's recreation of Doom, helping to generate the game's visuals in real-time as the AI 'plays' through the levels.

💡Rendering

Rendering in the context of video games and AI refers to the process of generating 2D images from 3D models and scenes. The video discusses how the AI project uses a modified version of Stable Diffusion for rendering, which is crucial for creating the visual output that mimics the original Doom game as the AI 'plays' it.

💡Training Data

Training data is the information used to train a machine learning model. In the video, the AI is trained on 900 million frames of Doom gameplay, which serves as the training data. This vast amount of data allows the AI to learn the patterns and mechanics of the game, enabling it to recreate the game's environment and gameplay.

💡Game Logic

Game logic refers to the rules and mechanics that govern how a video game functions. The video explains that while the AI can recreate the visuals and some aspects of gameplay similar to Doom, it does not truly understand or implement game logic like a traditional game engine would. The AI is essentially mimicking what it has learned from the training data rather than having a conceptual understanding of the game's rules.

💡Real-time

Real-time in computing refers to the ability of a system to respond to inputs within a short and predictable time frame. The video highlights that the AI's recreation of Doom is capable of running at over 20 frames per second in real-time, which is significant as it demonstrates the system's ability to process and render game visuals quickly enough to simulate a playable game experience.

💡Memory Model

In the context of the video, the memory model refers to the AI's ability to remember and learn from previous game states. The AI can only remember the last few seconds of gameplay, which limits its ability to create permanent changes in the game world. This is a limitation of the current system, as it does not have a long-term memory to retain the state of the game over extended periods.

💡VizDoom

VizDoom is a research platform for developing AI bots that can play Doom using visual information. It is mentioned in the video as the tool used to train the AI agents that provided the training data for the AI's recreation of Doom. VizDoom allows for reinforcement learning from raw visual information, which is essential for training AI in the context of video games.

Highlights

Google's research project, GamenGen, recreates Doom using AI.

The AI learns gameplay aspects such as opening doors and needing keys.

The AI can only remember the last few seconds of gameplay.

The AI recreates Doom by mimicking the original game's mechanics.

The AI has been trained on 900 million frames of Doom gameplay.

GamenGen runs at over 20 frames per second during gameplay.

The AI does not understand concepts like hit scanning or inventory.

The AI's rendering is done using a modified version of stable diffusion.

GamenGen is not a replacement for traditional game engines.

The AI's limitations include a lack of permanence and understanding of game mechanics.

The AI's training data was captured at a low resolution.

The AI uses previous frames to predict and render the next frame.

GamenGen's technology could potentially reduce game development costs.

The project aims to make game development more accessible through AI.

The AI's understanding of the game is based on模仿 rather than true comprehension.

The AI's limitations include memory constraints and the need for more training data.

The technology behind GamenGen is based on vizDoom, an AI research platform.

The project's future work includes testing the AI on other games and interactive software.

GamenGen's potential is seen as a step towards a new paradigm in game development.

The AI's ability to recreate Doom is impressive but not indicative of replacing game developers.