# SIMA 2
SIMA 2 is a generalist embodied agent developed by [Google DeepMind](https://www.edgepedia.org/google-deepmind), announced on November 13, 2025, that perceives 3D virtual worlds through raw video and acts through a keyboard-and-mouse interface while reasoning about goals, conversing with users, and accepting instructions in language or images.[1](https://arxiv.org/html/2512.04797v1) • [2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) It builds on SIMA ("scalable instructable multiworld agent"), an instruction-following game agent that debuted around mid-2024; where SIMA 1 executed simple, step-by-step language commands, SIMA 2 integrates DeepMind's Gemini models so that it can reason about high-level goals and describe its intended actions.[3](https://siliconangle.com/2025/11/13/google-deepminds-sima-2-agent-learns-think-act-inside-virtual-worlds/) • [4](https://www.infoq.com/news/2025/12/sima-2-gemini-agent/) The published benchmark results are vendor-reported, from DeepMind's technical report and blog post.[1](https://arxiv.org/html/2512.04797v1) • [2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/)
| Key fact | Detail |
| --- | --- |
| Announced | November 13, 2025, by Google DeepMind[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) |
| Backbone | Finetuned Gemini Flash-Lite model[1](https://arxiv.org/html/2512.04797v1) |
| Perception and action | 720p RGB frames in; 96 keyboard keys, mouse clicks, and discretized mouse movements out[1](https://arxiv.org/html/2512.04797v1) |
| Training environments | Three DeepMind research worlds plus seven commercial games[1](https://arxiv.org/html/2512.04797v1) |
| Headline vendor result | Beats SIMA 1 by over 10% in each tested environment; completes tasks in 26 of 50 MineDojo task categories versus 2 for SIMA 1[1](https://arxiv.org/html/2512.04797v1) |
| Availability | Limited research preview for a small cohort of academics and game developers[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) |
| Documented limits | Struggles with very long-horizon tasks; short interaction memory; precise control remains an open challenge[1](https://arxiv.org/html/2512.04797v1) |
## What SIMA 2 is
DeepMind frames SIMA 2 as the point at which the SIMA line evolved "from an instruction-follower into an interactive gaming companion."[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) The agent can follow language instructions, reason about goals, converse with users, and self-improve in virtual 3D worlds, and it handles complex instructions given through both language and images, where SIMA 1 required step-by-step direction.[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) • [4](https://www.infoq.com/news/2025/12/sima-2-gemini-agent/) DeepMind also presents it as a step toward real-world robots, although it declined to give a robotics timeline at launch.[3](https://siliconangle.com/2025/11/13/google-deepminds-sima-2-agent-learns-think-act-inside-virtual-worlds/) • [5](https://techcrunch.com/2025/11/13/googles-sima-2-agent-uses-gemini-to-reason-and-act-in-virtual-worlds/)
## Architecture and training as published
According to the technical report (published on arXiv in December 2025), the core of SIMA 2 is a finetuned Gemini Flash-Lite model, trained with supervised fine-tuning on a mixture of gameplay data and Gemini pretraining (non-gameplay) data so that reasoning, dialogue, and vision capabilities are preserved.[1](https://arxiv.org/html/2512.04797v1) The agent's input is a stream of RGB video frames at 720p resolution, and its output is structured text that is deterministically parsed into keyboard-and-mouse commands spanning 96 standard keyboard keys, mouse clicks, and discretized relative (x, y) mouse movements; the agent receives no privileged environment information.[1](https://arxiv.org/html/2512.04797v1)
Training data combined human demonstration videos with language labels and Gemini-generated labels, which let the agent describe its intended actions and steps.[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) The technical report lists the training environments as the research environments Construction Lab, Playhouse, and WorldLab plus the commercial games [Goat Simulator](https://www.edgepedia.org/goat-simulator) 3, Hydroneer, No Man's Sky, Satisfactory, Space Engineers, Valheim, and Wobbly Life, with Space Engineers newly added since SIMA 1; evaluation extended to unseen games including [Minecraft](https://www.edgepedia.org/minecraft) and ASKA.[1](https://arxiv.org/html/2512.04797v1) MIT Technology Review reported the training set as footage of humans playing eight commercial games plus three DeepMind-created virtual worlds; the technical report's own list names seven commercial games, and this discrepancy is unresolved.[6](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) • [1](https://arxiv.org/html/2512.04797v1)
Two further mechanisms round out the published picture. Synthetically generated "bridge data" of dialogue and reasoning, produced by a Gemini model, teaches the agent how to relate high-level user instructions and dialogue with internal reasoning and low-level actions.[1](https://arxiv.org/html/2512.04797v1) And the system is designed for open-ended self-improvement: Gemini proposes tasks and provides estimated rewards that are added to a bank of self-generated experience, which feeds the training of subsequent generations, so the agent can learn new skills from scratch without human demonstrations.[1](https://arxiv.org/html/2512.04797v1) • [2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/)
## Benchmark results: vendor versus independent
The quantitative results below are vendor-reported, from DeepMind's technical report and blog post.[1](https://arxiv.org/html/2512.04797v1) • [2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/)
On ASKA and on a subset of the [MineDojo](https://www.edgepedia.org/minedojo) benchmark in Minecraft, DeepMind reports that SIMA 2 outperforms SIMA 1 by over 10% in each environment.[1](https://arxiv.org/html/2512.04797v1) The MineDojo comparison is the sharpest: SIMA 1 completed only two task types (harvest dirt and combat spider), while SIMA 2 completes tasks in 26 of 50 task categories, which the authors attribute to the general world knowledge inherited from Gemini.[1](https://arxiv.org/html/2512.04797v1)
Finetuning for gameplay also costs the base model something. DeepMind reports regressions versus base Gemini of -4.0% on LCB (code) with supervised fine-tuning and -8.4% with SFT plus reinforcement learning; -25.5% on AIME (math) with SFT and -15.4% with SFT+RL; and -16.3% on GPQA Diamond (STEM) with SFT and -19.5% with SFT+RL, which the authors characterize as minor.[1](https://arxiv.org/html/2512.04797v1)
The paper reports that SIMA 2 "substantially closes the gap with human performance" across a diverse portfolio of games, while cautioning against over-interpreting direct comparisons because failure modes differ: humans more often failed due to time constraints, agents due to suboptimal exploration.[1](https://arxiv.org/html/2512.04797v1) MIT Technology Review's reporting on the same launch states the agent is far worse than humans at mouse-and-keyboard control.[6](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) Both statements are consistent with the paper's own caveat.
## By the numbers
- **+10% per environment** over SIMA 1 on ASKA and the MineDojo subset (vendor-reported).[1](https://arxiv.org/html/2512.04797v1)
- **26 of 50** MineDojo task categories completed, versus 2 for SIMA 1 (vendor-reported).[1](https://arxiv.org/html/2512.04797v1)
- **Seven commercial training games** (Goat Simulator 3, Hydroneer, No Man's Sky, Satisfactory, Space Engineers, Valheim, Wobbly Life) plus three research worlds, per the technical report; MIT Technology Review counted eight commercial games.[1](https://arxiv.org/html/2512.04797v1) • [6](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/)
- **96 keys** in the emulated keyboard action space, with mouse clicks and discretized relative mouse movements.[1](https://arxiv.org/html/2512.04797v1)
- **720p** RGB video input resolution.[1](https://arxiv.org/html/2512.04797v1)
- **Up to -25.5%** regression on AIME versus base Gemini after supervised fine-tuning (vendor-reported).[1](https://arxiv.org/html/2512.04797v1)
## Generalisation claims and independent skepticism
DeepMind's central generalisation claim pairs SIMA 2 with [Genie](https://www.edgepedia.org/genie-interactive-world-model) 3, DeepMind's world model that generates real-time 3D worlds from a single image or text prompt. According to the company, when challenged to play in these newly generated worlds, SIMA 2 could sensibly orient itself, understand user instructions, and take meaningful actions toward goals despite never having seen such environments before.[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) The paper likewise reports generalization to previously unseen environments.[1](https://arxiv.org/html/2512.04797v1)
Independent expert comment is more guarded. Matthew Guzdial, a researcher at the [University of Alberta](https://www.edgepedia.org/university-of-alberta) working on game-playing AI, told MIT Technology Review that most games share similar keyboard-and-mouse controls, so cross-game transfer is unsurprising, and that a game with unusual input would likely defeat the agent: "If you put a game with weird input in front of it, I don't think it'd be able to perform well."
[6](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/) He also questioned transfer from games to real-world robots, noting that real-world camera visuals are much harder to understand than game visuals, which are designed with easily parsable visuals for human players.[6](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/)
## Availability, licensing and use
SIMA 2 was announced as a limited research preview, with early access given to a small cohort of academics and game developers in consultation with DeepMind's Responsible Development & Innovation Team.[2](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) DeepMind gave no timeline for release beyond the preview; team member Wang told [TechCrunch](https://www.edgepedia.org/techcrunch) the goal is to show the world what DeepMind has been working on and attract collaborations.[5](https://techcrunch.com/2025/11/13/googles-sima-2-agent-uses-gemini-to-reason-and-act-in-virtual-worlds/)
## References
1. [SIMA 2: A Generalist Embodied Agent for Virtual Worlds (arXiv technical report)](https://arxiv.org/html/2512.04797v1)
2. [SIMA 2: A Gemini-Powered AI Agent for 3D Virtual Worlds — Google DeepMind](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/)
3. [Google DeepMind's SIMA 2 agent learns to think and act inside virtual worlds — SiliconANGLE](https://siliconangle.com/2025/11/13/google-deepminds-sima-2-agent-learns-think-act-inside-virtual-worlds/)
4. [SIMA 2 Uses Gemini and Self-Improvement to Generalize across Unseen 3D and Photorealistic Worlds — InfoQ](https://www.infoq.com/news/2025/12/sima-2-gemini-agent/)
5. [Google's SIMA 2 agent uses Gemini to reason and act in virtual worlds — TechCrunch](https://techcrunch.com/2025/11/13/googles-sima-2-agent-uses-gemini-to-reason-and-act-in-virtual-worlds/)
6. [Google DeepMind is using Gemini to train agents inside Goat Simulator 3 — MIT Technology Review](https://www.technologyreview.com/2025/11/13/1127921/google-deepmind-is-using-gemini-to-train-agents-inside-goat-simulator-3/)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Foundation-model methods and training › Multimodal, embodied, and world-model methods*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license