Google DeepMind Unveils SIMA 2: AI Agent Capable Of Playing, Reasoning, And Learning In 3D Virtual Worlds

In Brief

Google DeepMind introduced SIMA 2 AI agent that can understand instructions, reason, and teach itself new skills in virtual environments, nearing human-level task completion.

![@needshave, cгенерируй, пожалуйста, изображение на основе этого промпта в течение 30 мин:

Summary: Google DeepMind introduced SIMA 2, a Gemini-powered AI agent that can understand instructions, reason, and teach itself new skills in virtual environments, doubling its predecessor’s performance and nearing human-level task completion.

Keywords: Google DeepMind, AI agent](https://img-cdn.gateio.im/webp-social/moments-938a77d2fcc6a4da1d9a4125abc0616a.webp)

AI arm of the technology company Google, Google DeepMind introduced SIMA 2, the latest version of its Scalable Instructable Multiworld Agent, marking a move toward more capable and general-purpose AI agents

Built on the advanced reasoning capabilities of Gemini models, the system expands beyond basic instruction-following in virtual environments and now functions as an interactive companion that can interpret goals, converse with users, and refine its performance over time

The first SIMA model learned hundreds of language-driven actions across commercial video games by observing screen input and operating with virtual controls rather than integrated game mechanics

SIMA 2 advances this approach by embedding Gemini as its core, enabling the agent to perform goal-directed reasoning, explain its intended actions, and execute more complex tasks within games. Trained on a combination of human demonstrations and Gemini-generated annotations, the agent has been tested across a broader set of games through partnerships with multiple developers. This update represents a significant step for embodied AI, combining perception, reasoning, and action within dynamic 3D environments.

The integration of Gemini has strengthened SIMA 2’s ability to generalize and operate reliably across unfamiliar contexts. The agent can now interpret more detailed and nuanced instructions and execute them successfully even in games it has not previously encountered, such as the Viking-themed title ASKA or MineDojo, a research version of Minecraft

Its capacity to apply learned concepts across different environments—for example, extending the idea of “mining” from one game to “harvesting” in another—forms a key component of broad generalization and brings its performance closer to that of a human player

In order to evaluate these capabilities, SIMA 2 was also tested within procedurally generated 3D worlds created by Genie 3, which produces new environments from text or image prompts. In these unfamiliar settings, the agent was still able to navigate effectively, interpret instructions, and work toward user-defined goals, showing a level of adaptability not previously observed in similar systems.

SIMA 2 Advances Self-Improving AI With New Capabilities In Generalization And Autonomous Learning

According to the company, one of SIMA 2’s most notable developments is its emerging ability to improve its own performance. During training, the agent has demonstrated that it can take on increasingly complex tasks through iterative trial-and-error combined with feedback from Gemini. After learning initially from human demonstrations, SIMA 2 is able to continue progressing in new games through autonomous play, gaining skills in unfamiliar environments without requiring additional human data. This experience can then be used to train subsequent, more capable versions of the AI agent, and the same self-improvement process has been applied successfully within Genie-generated environments, marking a meaningful advance toward training general agents across diverse, synthetic worlds. This cycle of continual refinement supports the longer-term aim of enabling agents to learn with minimal human guidance.

SIMA 2’s operation across a wide range of gaming environments provides an important testing ground for general intelligence, allowing it to acquire skills, practice reasoning, and learn continuously through self-directed action. Although the system represents a substantial step toward generalist, interactive, embodied intelligence, it retains clear research-stage limitations. The agent continues to struggle with complex, long-horizon tasks that require extended reasoning or repeated goal verification, and its memory remains short due to the need for low-latency interaction within a limited context window. Precision in fine-grained actions and visual understanding of complex 3D scenes also remains a broader challenge across the field.

The project demonstrates the potential of an action-oriented AI approach in which broad competency is supported by diverse training data and strong reasoning capabilities. SIMA 2 shows that these elements can be unified in a single generalist agent rather than isolated in separate specialized systems, and it provides a promising path toward future applications in robotics, as many of the skills learned in virtual settings—such as navigation, tool use, and collaborative task handling—translate into fundamental components for embodied AI.

SIMA 2 is designed as an interactive, human-centered research agent, and its development includes a clear focus on responsible practices, particularly concerning its self-improvement mechanisms. The team has collaborated with responsible innovation specialists throughout the project and is releasing SIMA 2 in a limited research preview, providing early access to selected academics and game developers. This phased approach allows for continued scrutiny, feedback, and interdisciplinary evaluation as the technology and its potential implications are further explored.