Google announces AI ``SIMA'' that can play nine games such as ``Goat Simulator 3'' and ``No Man's Sky'' just by giving instructions in human words



Google has announced an AI ``Scalable Instructable Multiworld Agent (SIMA)'' that can play various games according to human instructions in natural language.

SIMA generalist AI agent for 3D virtual environments - Google DeepMind
https://deepmind.google/discover/blog/sima-generalist-ai-agent-for-3d-virtual-environments/

Google's new gaming AI aims past “superhuman opponent” and at “obedient partner” | Ars Technica
https://arstechnica.com/ai/2024/03/google-deepmind-wants-its-sima-ai-model-to-be-the-ultimate-co-op-gaming-partner/

Google has created AI that can play Go , Atari , and StarCraft II with better skill than humans.

AI 'AlphaStar' who won a big victory over professional gamers is promoted to grand master, which is the top 0.2% of humanity in 'StarCraft 2' - GIGAZINE


by Marco Verch Professional Photographer and Speaker

The ``Scalable and Directable Multi-World Agent (SIMA)'' developed by Google differs from previous AI in that its goal is to play according to human instructions, rather than to beat high scores or beat humans. It was made by

Google engineer Tim Hurley said, ``SIMA wasn't trained to win games, it was trained to do what you were told. It just learned to play one game.'' 'But while it's a technical feat for an AI system, learning to follow instructions in a variety of games could unlock AI agents that are more useful in any environment.'

To develop SIMA, Google worked with eight game studios to teach SIMA how to play nine games, including Goat Simulator 3, Hydroneer, No Man's Sky, Satisfactory, and Valheim. .

In training the AI, we first had it learn the basic movements of objects in an environment built with Unity, and then learned it through human play. In human game play, one player plays the game while the other instructs the next action, or one player plays the game freely and then moves on. The participants were asked to explain how their behavior led to their actions in the game.

Having learned how to play the game in this way, SIMA can complete simple tasks within about 10 seconds by receiving verbal instructions such as 'shoot an asteroid,' 'gather resources,' and 'jump over a fence.' Now you can do about 600 types.



What's notable about this study is that SIMA, which was trained on all nine games, outperformed agents trained exclusively on one particular game. Specifically, when the training data included another game, the performance was 67% higher than when it was not included.

SIMA's development team will continue to undergo more extensive training to develop an even more general-purpose and versatile AI, and in the future, it will be able to handle complex tasks such as ``procuring materials and building a camp.'' We want to be able to complete tasks.

'Ultimately, our research aims to build more general AI systems and agents that can understand and safely perform a wide range of tasks to help people not only online but also in the real world. ” said the SIMA team.

Related Posts:

in Software,   Video,   Game, Posted by log1l_ks