The next generation of video games will have artificial intelligence built in, and that moment won't be long. Already Google DeepMind, a subsidiary of Google, is preparing SIMA.
Google SIMA (Scalable Instructable Multiworld Agent) is a productive artificial intelligence product developed by Google DeepMind and trained in virtual 3D environments to perform basic game tasks following physics instructions languages given by the user.
According to Google, the ultimate goal is to make SIMA assistants more advanced so that they can safely perform complex tasks within video games and potentially the real world.
How does SIMA work?
SIMA AI works by taking two simple inputs: image processing (game) and real-time language commands provided by the player. This Google AI does not need permission to perform a range of tasks, such as modifying source code, APIs or special privileges to act as a game assistant.
Η τρέχουσα έκδοση του SIMA περιέχει 600 βασικές δεξιότητες, που εκτείνονται σε πλοήγηση (π.χ. "στροφή αριστερά"), αλληλεπίδραση αντικειμένων ("ανέβα στη σκάλα") και χρήση μενού ("άνοιγμα χάρτη"). According to Google have trained SIMA to perform simple tasks that can be completed in about 10 seconds.
At its core, SIMA has two models: one for language mapping-image και το άλλο για πρόβλεψη βίντεο. Το προηγούμενο μοντέλο γλώσσας-εικόνας βοηθά κυρίως το SIMA στην κατανόηση των εντολών φυσικής γλώσσας και της σχέσης τους με το οπτικό content of the screen. The video model helps SIMA predict future events so it can pre-plan its actions in games.
With both of these AI models, SIMA combines visual observations and language instructions to perform short game tasks (
Having your SIMA helper in the game will be almost like having a human helper with you. However, this AI gaming tool is not about “achieving high game scores”. The intention of the technology is to slowly introduce artificial intelligence into games and take the next step which is its entry into virtual environments.
How was SIMA trained?
SIMA was trained on nine different commercial video games, including No Man's Sky by Hello Games, Teardown by Tuxedo Labs, and Valheim by Iron Gate.
Developers choose to train SIMA on games open word and sandbox to help the AI learn a wide range of elementary game skills such as navigation, shooting, digging, driving, crafting, etc.
SIMA's training focused on first- or third-person gameplay, while avoiding games with extreme violence (for now).
A vital part of SIMA's training is behavioral cloning, where AI assistants learn by observing data generated by expert players. This dataset included gameplay videos, walkthroughs, commentary, and more.
Using SIMA as an AI assistant in games
SIMA is currently under research and its use as an AI assistant is not ready, as it is still at a very rudimentary stage, e.g. moving, picking up tools, getting into a vehicle, etc.
In fact, it is not yet available for public beta testing, and there is no option to deploy Google SIMA at the individual or commercial gaming level.
However, SIMA integration would require minimal technical skills when released. This gaming AI takes only two inputs: visual images and language instructions, with no root-level access.
Can SIMA play games with you?
Yes, in theory, SIMA will be able to play games alongside you when it goes on sale. While it may not perform at the same level as human players experienced in gaming, players should benefit from SIMA's abilities to perform basic game tasks or repeat routines whenever necessary.
Future prospects of SIMA
The developers evaluated SIMA on 1.485 unique game tasks across nine skill categories, including movement, simple navigation, resource gathering, object manipulation, and more.
Compared to humans, SIMA performed quite well. For example, from the game No Man's Sky, SIMA succeeded 34% of the time, while human players succeeded 60% for the same set of tasks.
The training also confirms that SIMA can perform well as a generalized AI assistant. This means developers don't have to train SIMA on every other game.
This finding was confirmed in the developers' technical report, which states that AI assistants trained on multiple games outperformed assistants trained on a single game. Even in games they hadn't seen.
Consequently, this may translate into a future where Google may release a single, universal SIMA assistant for all game genres, or at least game-specific SIMA assistants. In such cases, it can be a subscription-based service that users can simply integrate into their games.
In addition, Google can work with game studios to combine SIMA to release special AI-assisted versions of popular games.
However, Google conveyed the intention to develop artificial intelligence systems that are useful not only in virtual 3D environments such as games, but also in real, natural life.
