Google Launches SIMA 2, An AI Agent That Learns Virtual 3D Worlds And Assists Gamers
The SIMA 2 can follow human language instructions in video games, think and reason in-game queries, and learn new skills.
Published : November 17, 2025 at 6:17 PM IST
Hyderabad: Google’s AI research wing, DeepMind, has introduced Scalable Instructable Multiworld Agent (SIMA) 2. An upgraded version of the SIMA, the new artificial intelligence (AI) agent comes with several enhancements over its predecessor. It is equipped with Gemini AI models that enable the agent to execute a certain action, provide reasoning for it, and also interact with the user using a text interface.
The main function of SIMA 2 is to be a versatile and adaptable AI companion, which can assist and collaborate with humans in various tasks and environments.
What is SIMA 2?
The SIMA 2 is a generalist AI agent that can understand and interact with virtual worlds, follow instructions, think and reason, and enhance itself over time. According to Google, SIMA was a huge leap in teaching AI how to translate language into meaningful actions in the 3D worlds of video games.
SIMA 2 is our most capable AI agent for virtual 3D worlds. 👾🌐
— Google DeepMind (@GoogleDeepMind) November 13, 2025
Powered by Gemini, it goes beyond following basic instructions to think, understand, and take actions in interactive environments – meaning you can talk to it through text, voice, or even images. Here’s how 🧵 pic.twitter.com/DuVWGJXW7W
What’s new: The new AI agent, with the help of Gemini AI models, has now turned into an interactive gaming companion rather than an instruction-following agent, which was the case in SIMA.
It can follow human language instructions in virtual worlds, think and reason about the game’s goals and tasks. SIMA 2 can interact with users and provide answers to their queries related to the game. Along with this, the new AI agent comes with advanced follow instructions and reasoning, and teaches itself new skills in the video game. It can understand multimodal prompts given by the user, different languages and even emojis. DeepMind has not mentioned the number of languages the SIMA 2 agent can understand.
What can SIMA 2 do?
DeepMind states that SIMA 2 can analyse a player’s actions and learn the steps required to complete a certain task. The AI agent receives visual information from a three-dimensional game and understands a user-defined objective like “Fly to a red flower”, “Obtain cobblestone with an iron pickaxe”, “Extinguish the campfire”, and more. It breaks user-prompted goals into smaller actions and performs them using inputs similar to a keyboard and mouse. This approach allows SIMA 2 to execute user instructions based on what happens on screen.
One of the highlighted features of the AI agent is its enhanced performance in unfamiliar games. The SIMA 2 was tested on video games such as Minedojo — a research-focused adaptation of Minecraft, and ASKA — a Viking-themed survival game. DeepMind states that the latest AI agent outperformed SIMA by showcasing better adaptability and higher success rates in game tasks.
Another key feature of the AI agent is its ability to apply concepts learned in one game to another, enabling more efficient learning across diverse game environments. Whenever SIMA 2 learns a new skill or movement in a new environment, the experience is recorded and put into the training process.
DeepMind has mentioned that SIMA 2 struggles with long-term memory, complex multi-step reasoning and goal verification, and executing precise low-level actions.
The SIMA 2 showcases significant steps in the Artificial General Intelligence (AGI), which will help to shape the future of robotics and AI-embodiment.
