This AI factor can radically change gameplay, because it learns completely differently

We have associated artificial intelligence (AI) with the video game medium since its early days. After all, he had already entered Pong The computer-controlled opponent pushed the pixels to our side. However, this term is usually used somewhat loosely. Because it is very rare that what we experience in a video game in terms of the behavior of opponents or NPCs in general is actually artificial intelligence.

Rather, so-called scripts are executed. If-then scenarios with many branches, for example an opponent's reaction to being shot in a conversation Call of duty Make it look almost realistic. But Google is offering this now Team research on artificial intelligence Deepmind This can really change the way you experience video games.

“General AI agent for 3D virtual environments”

Sima is what Google calls an AI agent. This means “scalable and teachable multi-world agent”. Pretty cool title, but it pretty much shows what Sima can do. Or at least they should be able to. What distinguishes it is that it does not require access to a specific game code or any special interfaces. Instead, for the agent to learn, it only needs images displayed on the screen and verbal instructions provided by the user. This means that this approach is not only applicable to a single game, but is scalable on a large scale as well.

For training purposes, Google scientists worked with eight game studios. Including the coffee spot, which many Valheim You should know, and Hello Games, which is specifically for No Sk many well known. “Each game in the Sima portfolio opens up a new interactive world, including a set of skills to learn. From simple navigation and using the menu to extracting resources, flying on a spaceship or crafting a helmet.

In the first step, pairs of human players were recorded, while one part observed and gave instructions, while the other part actively played the game. We also let the players play freely and then look back at what they did in order to record instructions that would help them with this. “It led to their actions in the game,” she also says.

This is how an AI agent works

So Sima learned not through the game code, but through the actions of the players, based on what the screen showed and the verbal instructions that were given. And this is exactly how the agent behaves in the game. It's supposed to do what you ask it to do – while also reacting to what's happening around it in the game.

The agent is currently said to have 600 basic skills, such as 'turn left', 'climb the ladder' or 'open the map' – and is said to have already outpaced AI models that were only trained on a specific video. Game. “Furthermore, the agent trained in all but one game performed, on average, in this unseen game almost as well as the agent trained specifically for it,” Google said in the text. This shows that Sima can work even in an unfamiliar environment – and can therefore do more than just what he has learned.

However, language instructions are always important. Without this, Sima would act “appropriately” but “without purpose.” The agent will then collect resources because he has learned that this is a common activity in the game, but he will not follow a specific path he has been directed to do.

What will this change?

First, it's important to note that Sima is still in development. But: “We want our future agents to complete missions that require extensive strategic planning and multiple subtasks, such as finding resources and building a camp.” So it's safe to assume that Sima, or a similar AI agent, will indeed appear in a commercial video game in the not-so-distant future.

This could radically change the way video games work. As we mentioned at the beginning, games will no longer rely on scripts that always work according to a similar scheme. Characters in a video game will not run on relatively fixed paths. Instead, based on their training, they will move more or less freely in the game – depending on the instructions you give them as players.

But the game itself can also be changed dramatically through this. When you no longer control the game by physical input, but by speaking to the game – giving it instructions which are then carried out according to your wishes.

Initially, something like this might have to happen within a more limited framework in terms of the instructions you can provide. The more choices there are in the game, the more likely errors are to occur. And probably no one wants a game where NPCs are completely free to do what they want – because then the game will be turned upside down in a way you don't want it to be.

However, Sima's current capabilities already show that AI agents that interact one-on-one with players is a not-too-distant future. Especially when it is not developed for individual games, but can also be used as a scalable model by small studios. Because once this AI is trained, it can make almost any 3D environment unsafe.

This is how developers interact with it

We've already spoken with some developers about the new AI agents. Of course, no one has been able to try it yet, but the initial feedback is very positive. “This sounds exciting,” says Paul Lawitsky, programmer and game designer at Chasing Carrots. “This could be a good entry point to have more interesting in-game characters playing with or against the player in a more solution-oriented way,” he explains. This development could also provide a boost to accessibility.

“People who can't use a keyboard and mouse or gamepad can have the opportunity to experience games in a different way,” Lawitsky said. Eventually, they can then give instructions by voice without having to make inputs with a button or analog stick.

6 examples of how AI is already changing video games: