Need a co-player in video games? Google's SIMA 2 shows how AI could help

SIMA 2 builds on Google's earlier game-trained agent with Gemini-powered reasoning, letting it follow complex instructions, adapt to new environments, and collaborate like a real in-game companion

Google has introduced SIMA 2, the next version of its AI agent designed to act, learn, and reason inside 3D virtual worlds. Unlike the first-gen SIMA, which could only follow simple instructions like “turn left” or “open the map,” the new SIMA 2 behaves more like a co-player — one that can understand what you want, talk back, plan its next move, and even improve itself over time.

 

SIMA 2 is still an early research project, not a consumer tool, but it gives a clear look at how AI could eventually assist players inside games.

What is Google’s SIMA 2?

SIMA 2 (short for Scalable Instructable Multiworld Agent) is a virtual AI assistant that can play PC games the same way humans do — by looking at the game screen and using a virtual keyboard and mouse. Google DeepMind trained it across several commercial titles such as No Man’s Sky, Valheim, Goat Simulator 3, and more, without giving it access to any hidden game data or cheat-like controls.

 

Where SIMA 1 could execute more than 600 basic actions, SIMA 2 takes a far bigger leap:

• It understands natural-language instructions
• It reasons about the goal instead of blindly following commands
• It can explain what it plans to do
• It can handle new games it has never seen before
• It can learn new skills by playing on its own

Google says interacting with SIMA 2 feels less like issuing commands to an AI and more like collaborating with an in-game companion that understands the broader task.

How is SIMA 2 different from existing gaming AIs or NPCs?

Today’s video-game AIs or NPCs (non-player characters) operate only inside their scripted worlds. They follow pre-written patterns and cannot adapt to completely new situations or games.

 

SIMA 2 is different in a few key ways:

 

It uses Gemini’s reasoning abilities: Google has embedded its Gemini model inside SIMA 2, allowing the agent to think through tasks, set goals, and choose actions logically. It’s not just following instructions, it’s interpreting them.

 

Example: When asked to “find a campfire,” SIMA 1 often failed. SIMA 2 can scan the environment, understand what a campfire looks like, and figure out where it is, even in a game it has never played.

It generalises between games: If SIMA 2 learns what “mining” looks like in one game, it can apply that idea to “harvesting” in another. This is closer to how humans think and is a major step toward general AI.

 

It understands sketches, images, and emojis: Players can draw a rough shape on the screen, send an emoji prompt, or speak in different languages. SIMA 2 can interpret all of it.

 

It improves itself without human help: Once it learns the basics from human demonstrations, SIMA 2 can continue training by playing on its own. It tries tasks, fails, learns, and gets better.

 

It works in AI-generated worlds too: Google tested SIMA 2 in new 3D worlds created by its Genie model (which generates game worlds from text or images). SIMA 2 could orient itself and take meaningful actions, despite these being entirely new environments.  ALSO READ | After OnePlus 15, iQOO and Realme to launch Snapdragon 8 Elite Gen 5 phones

What could SIMA 2 lead to in the future?

While Google is not positioning SIMA 2 as a gamer tool you can download today, the technology hints at several future possibilities:

• AI co-players in games: Imagine an AI teammate that helps you explore open worlds, complete quests, or handle resource grinding, without needing any game integration or modding.
• Smarter NPC behaviour: Game developers could use SIMA-like agents to create NPCs that react more naturally and adapt to player choices.
• Training ground for real-world robots: Google says video games are an ideal testbed for embodied AI. Skills like navigation, tool use, planning, and collaboration can eventually translate into robotics.
• Universal language-to-action assistants: SIMA 2’s ability to convert natural language into actions could power future assistants that interact with digital environments more freely — not just games.

When will SIMA 2 be available?

SIMA 2 is being released only as a limited research preview for select academics and game developers. Google is treating this as a controlled experiment, citing the need for responsible development, especially because the agent can self-learn and improve without human supervision.