Do you think artificial intelligence can be your gaming partner? Discover what Google DeepMind has achieved!

Has Google DeepMind succeeded in creating an artificial intelligence capable of playing video games with you, but also understanding and responding to your verbal instructions?

Here are the latest advances from DeepMind in the gaming AI field, highlighting their latest fascinating project: an AI model capable of learning to play different 3D games as a human would.

Playing multiple games like a human

Traditionally, AI models specialized in games are limited to a single game and always play to win. But DeepMind has chosen a different approach: training an AI model to play multiple 3D games, while teaching it to understand and act on verbal instructions.

DeepMind’s SIMA (Multi-World Instructable and Extensible Agent) is an artificial intelligence developed by Google DeepMind researchers.

It is an AI model designed to learn to play various 3D games in a manner similar to a human, while being capable of understanding and acting on verbal instructions.

Unlike some AI models specialized in a single game, SIMA is trained on multiple games, enabling it to generalize its skills and play games it has not yet encountered. The goal is to create a more natural and cooperative virtual gaming companion, capable of interacting with players in a more intuitive way.

SIMA is trained using many hours of videos showing humans playing different games, without access to internal code or specific game rules. Based on this data, the AI learns to associate actions, objects, and interactions with visual representations, allowing it to understand and reproduce behaviors similar to those of human players.

DeepMind's long-term goal is to develop more general and flexible AIs, capable of adapting to a variety of tasks and situations in the real world. SIMA represents a significant step forward in this direction, opening new perspectives for the use of AI in gaming and other application areas.

Learning without access to game code

DeepMind's AI, called SIMA (Multi-World Instructable and Extensible Agent), was trained without access to the internal code or rules of the games. Instead, it was fed many hours of videos showing humans playing different games, enabling it to associate actions, objects, and interactions with visual representations.

The universality of skills

A major question was whether an AI trained on a set of games could play other games it hadn't seen before. The results showed that yes, with some limitations. AI agents trained on multiple games performed better on games they hadn't yet encountered, but some games have unique mechanics that can pose challenges.

Training challenges

Although games have their own jargon, there are only a limited number of "verbs" that are truly meaningful for the AI. This means that the AI must learn from training data to identify and interpret these actions in a general way.

Comparison with other approaches

DeepMind uses a imitation learning approach based on human behavior, which differs from the simulator-based approach used in other cases. This method allows the AI to learn to perform a wide variety of tasks described in open text, without being limited by a strict reward system.

The future of AI in gaming

DeepMind's researchers' ambition goes beyond merely creating agents capable of playing games. They aim to develop more natural and cooperative gaming companions, who could one day accompany you in your gaming adventures.

The progress made by Google DeepMind in the field of gaming AI opens exciting new perspectives. Although many challenges remain, it is undeniable that we are getting closer and closer to creating virtual gaming partners that can rival the complexity and flexibility of the human mind.