SIMA 2

A Gemini-based Google DeepMind agent that follows instructions, reasons, and self-improves across many different 3D game worlds.

SIMA 2 was released by Google DeepMind in November 2025 as a limited research preview, with a technical report published in December. Its predecessor, SIMA (Scalable Instructable Multiworld Agent, released 2024), could carry out more than 600 simple skill instructions, like 'turn left' or 'climb the ladder,' across a range of commercial 3D games. SIMA 2 is built around a Gemini model and goes beyond following instructions: it can understand a high-level goal, hold a conversation with the user, explain what it plans to do, and accept instructions in the form of images, sketches, or even emoji. It can also self-improve: Gemini sets it tasks and scores its performance, and the agent trains its next version on the experience it generates by playing, no longer relying on human demonstrations. It can also orient itself and act on instructions in entirely new worlds generated in real time by Genie 3. DeepMind views SIMA 2 as a step toward a general embodied agent that will eventually be used on real robots.

ExampleTold to 'go to the house that's the color of a ripe tomato,' SIMA 2 first reasons that a ripe tomato is red, then goes looking for the red house.

Also called

Scalable Instructable Multiworld Agent 2

Related

Genie 3 · Embodied Agent · Self-improvement · Google Gemini · Google DeepMind · Instruction Following

Sources

SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds (Google DeepMind blog)
SIMA 2: A Generalist Embodied Agent for Virtual Worlds (arXiv 2512.04797)

As of

2025-12