Google Unveils SIMA: Artificial Intelligence That Can Play Video Games Like a Human

New from Google in the field of artificial intelligence Idleness, an agent capable of playing video games like a human. Mountain View staff have worked with several studios and taught technology using games such as Goat Simulator 3, Pluck And No Man’s Skyamong the others.

Google DeepMind researchers explained that the goal of SIMA is not to create an AI that can score high or win a game. On the contrary, the idea is that it can learn to interpret instructions in natural language to achieve goalswhile recognizing the three-dimensional environment presented in the game and anticipating what might happen next.

The company clarified that the new technology is part of a research project. Therefore, it does not aim to become a replacement for AI that is already used in various video games, for example, to enhance the capabilities of NPCs. In its current state, it will function as a sort of companion for the human player, capable of responding to specific requests to complete various tasks.

Google noted that video game training from Tuxedo Labs, Coffee Stain Studios and Hello Games, among others, enabled SIMA to acquire about 600 basic skills. These range from using in-game menus to navigating scenarios and interacting with objects found within them.

These skills focus on simple tasks that can be completed in about 10 seconds. For example, ask the AI ​​controlling the character to climb stairs, drive a car, open a map, or head in a certain direction.

SIMA, Google’s artificial intelligence that can play video games like a human

The learning method that Google used with SIMA is very interesting. Initially, the company studied the experiences of pairs of players, one of whom directed the game itself, and the other gave him instructions on what to do next. He also recorded each subject’s individual gaming sessions and analyzed their actions to generate instructions that would produce the same results.

In addition to this, DeepMind experts have developed a framework in Unity dedicated exclusively to artificial intelligence for learning to manipulate objects. For this purpose, a system for constructing structures using blocks was given.

The interesting thing about SIMA, as Google explained, is that it doesn’t require any special APIs to train. You should also not access video game source code.. Here’s how the Californians explain it:

“[El agente] only two input methods are required: images on the screen and simple natural language instructions provided by the user. SIMA uses keyboard and mouse commands to control the game’s main character and carry out those instructions. “This simple interface is what people use, meaning SIMA can potentially interact with any virtual environment.”

Google DeepMind.

Google believes that as SIMA becomes familiar with more learning worlds, it will become more versatile. This way you can understand and carry out orders and tasks of greater complexity. Since this is a research project, it is unknown if this technology will reach the general public at any point.

Californians are eager to make these types of AI agents useful beyond their specialized training in video games. “Ultimately, our research is focused on more general AI systems and agents that can safely understand and perform a wide range of tasks in ways that are useful to people online and in the real world,” they pointed out.