Recently, researchers have developed an AI model named DIAMOND (Diffusion for World Modelling), which is capable of simulating the renowned computer game Counter-Strike: Global Offensive (CS:GO) within neural networks.
This model operates on an Nvidia RTX3090 graphics card, achieving a speed of 10 frames per second. Although this frame rate is not particularly high, it remains a noteworthy achievement in the field of AI simulation.
DIAMOND's training data consists of merely 87 hours of CS:GO gameplay recordings, which is only 0.5% of the data required for similar projects like GameNGen. Despite the small data volume, the model can still simulate impressive in-game scenarios.
DIAMOND initially demonstrated its capabilities on Atari games, employing a Transformer-based approach, treating player movements as "tokens," akin to words in a sentence. By predicting these tokens, the model learns to anticipate the player's next move based on previous actions.
Researcher Eloi Alonso showcased the model's capabilities on Twitter, where videos depicted players interacting with the simulated CS:GO environment via keyboard and mouse. The simulation included complex elements such as player interactions, weapon mechanics, and environmental physics. However, DIAMOND still has some glaring flaws.
For instance, players can jump infinitely because the model does not account for gravity or collision detection in the Source engine. Additionally, once a player deviates from commonly used paths in the training data, the simulation completely collapses.
Researchers believe that as the amount of data and computational power increases, the model's performance will further improve. They also envision the potential for developing AI models capable of navigating complex real-world environments in the future.
It is worth noting that DIAMOND's CS:GO simulation was inspired by the GameNGen system developed jointly by Google Research, Google DeepMind, and Tel Aviv University, which can fully simulate parts of the classic game DOOM at 20 frames per second on a single Google TPU chip.
For developers interested in AI, the DIAMOND model is now open-source on GitHub, inviting further exploration.
Project entry: https://diamond-wm.github.io/
Key points:
- 🎮 The AI model DIAMOND developed by researchers can simulate CS:GO, running on an Nvidia RTX3090 at 10 frames per second.
- 📊 The model was trained using only 87 hours of game data, yet it can still simulate complex game scenarios despite the small data volume.
- 🐛 DIAMOND has some severe limitations and bugs, but researchers believe future improvements through increased data and computational power are possible.