DeepMind's SIMA 2 brings Gemini into 3D video games

Playing video games with an AI teammate usually feels like dragging a rigid script along—give it an instruction outside its playbook and it runs directly into a wall. Google DeepMind wants to fix that.

DeepMind introduced SIMA 2, a research preview of its next-generation generalist agent for 3D virtual worlds. While SIMA 1 arrived last year with 600 basic skills like climbing ladders and opening maps, SIMA 2 embeds Google's Gemini 2.5 Flash-lite model as its core reasoning engine. Instead of merely mapping buttons to commands, the agent now talks back, explains its plan, and learns on its own.

Reasoning through emoji and alien planets

SIMA 1 was an impressive demo, but its success rate on complex tasks lingered at 31%, miles behind the 71% benchmark of human players. With Gemini under the hood, SIMA 2 effectively doubled that performance.

Because it actually reasons now, it handles commonsense nuance. DeepMind researchers demoed asking the agent to walk toward the house matching the color of a ripe tomato; SIMA 2 reasoned internally that ripe tomatoes are red, scanned the 3D scene, and headed to the red house. It even handles multimodal shortcuts: drop the emoji "🪓🌲" into chat, and it fetches an axe to chop wood.

More importantly, SIMA 2 generalizes across games it has never seen, such as the Viking survival sim ASKA or Minecraft research environment MineDojo. It can take the concept of mining learned in one title and adapt it to harvesting in another. DeepMind pushed this further by pairing SIMA 2 with Genie 3, its world model that generates real-time 3D spaces from text prompts. Dropped into a newly imagined world, SIMA 2 navigated around benches, trees, and butterflies without missing a beat.

Self-improvement without human gameplay

The biggest architectural shift is how SIMA 2 learns. The first version relied entirely on hundreds of hours of recorded human gameplay. SIMA 2 uses that human data only as an initial baseline.

In new environments, DeepMind pairs the agent with two other models: another Gemini model that generates custom tasks, and a reward model that scores the attempts. SIMA 2 teaches itself by trial and error, turning its own simulated experience into training data for future iterations. This self-directed loop drastically cuts down the need for expensive human demonstrations.

DeepMind notes that this remains a limited research preview for select academics and game partners. Long-horizon planning is still brittle, and to keep interaction latency low, the agent runs on a relatively short memory window.

Why it matters

DeepMind is playing video games to solve the physical world. While building autonomous humanoids requires precise motor control, the much harder bottleneck is teaching an AI how to perceive an environment, reason about ambiguous goals, and plan multi-step actions.

Training robots in the physical world is painfully slow, prone to hardware damage, and choked by scarce real-world data. By turning commercial video games and procedurally generated Genie worlds into massive training grounds, DeepMind has built a synthetic gym for embodied AI. The agent learns tool use, spatial orientation, and self-directed trial and error at scale.

For gamers and game developers, this offers a peek at the death of scripted NPCs. Instead of hard-coded behavioral trees, virtual companions could soon behave like real players who understand natural banter. But for the broader industry, SIMA 2 is proof that large multimodal models can leap off the chat interface and pilot a body through complex spaces—a crucial stepping stone toward real-world robotics.