Tech giant Google’s artificial intelligence division, Google DeepMind, has introduced SIMA, an AI agent that has the potential to play video games like a professional human gamer.
Updated October 2026. This March 2024 report now corrects a few details against DeepMind’s announcement and adds a section on SIMA 2, the Gemini-powered successor DeepMind introduced in November 2025.
When it was announced, the AI model was a research project designed to follow instructions inside games the way a human player might, rather than to beat high scores. SIMA, which stands for Scalable Instructable Multiworld Agent, is a generalist agent intended to work across many different 3D game worlds instead of being built for a single game.
SIMA does not aim to become a replacement for the existing gaming AIs, and it will just gel in as a virtual co-op player. This learning AI would be able to understand verbal instructions while interpreting the 3D and virtual worlds as well. Rather than becoming a hard-coded AI bot, SIMA will provide its users with a more human-like gaming experience.
Google DeepMind stated in a blog post,
This research marks the first time an agent has demonstrated it can understand a broad range of gaming worlds and follow natural-language instructions to carry out tasks within them, as a human might. This work isn’t about achieving high game scores.
Notably, Google DeepMind has a long history in AI and games, from its early work on Atari games to AlphaGo, which mastered the board game Go, and AlphaStar, which plays StarCraft II at human grandmaster level. With SIMA, it shifted its focus from individual games towards a general, instructable game-playing agent.
As per the blog post, SIMA was trained to perform simple tasks that can be completed within about 10 seconds, and it was evaluated across 600 basic skills, such as turning left, climbing a ladder and opening the map. It plays using a keyboard-and-mouse interface, as a person would, by looking at the screen.
To train and test SIMA, Google DeepMind collaborated with eight game studios on nine different video games, including Hello Games (No Man’s Sky), Tuxedo Labs and Saber Interactive (Teardown), Coffee Stain (Valheim, Satisfactory and Goat Simulator 3), Keen Software House (Space Engineers) and Strange Loop Games (Eco).
Update: SIMA 2 arrives with Gemini
On November 13, 2025, Google DeepMind introduced SIMA 2, which embeds a Gemini model at the agent’s core. According to DeepMind, this moves SIMA from an instruction-follower towards an interactive gaming companion: SIMA 2 can reason about a user’s high-level goals, explain what it intends to do, answer questions and converse with users while it plays.
DeepMind says SIMA 2 generalizes much better than the original, successfully carrying out tasks in games it was never trained on, such as the Viking survival game ASKA and MineDojo, a research version of Minecraft. It can also follow sketches, prompts in different languages and even emojis, and it was able to act sensibly inside new 3D worlds generated by DeepMind’s Genie 3 world model. SIMA 2 can improve itself through self-directed play, using Gemini-based feedback rather than additional human gameplay data.
DeepMind describes SIMA 2 as a research project with clear limits: it still struggles with very long, complex tasks, has a relatively short memory, and finds precise low-level keyboard and mouse control challenging. It was released as a limited research preview for a small group of academics and game developers, not as a consumer product. For more on Google’s Gemini models, read The AI War – ChatGPT Vs. Gemini AI: Which One Makes the Wiser Choice?
