TLDR: Google DeepMind has introduced SIMA 2, a new AI agent designed to learn, adapt, and operate in virtual 3D environments with human-like reasoning. Powered by the Gemini AI model, SIMA 2 demonstrates enhanced capabilities in understanding high-level goals, explaining its plans, and collaborating within games. It achieved a 65% task completion rate, a significant improvement over its predecessor, SIMA 1’s 31%. The agent can also navigate and act in newly generated 3D worlds created by DeepMind’s Genie 3 project, showcasing its ability to transfer learned concepts across diverse tasks. This development is considered a crucial step towards Artificial General Intelligence (AGI) and has implications for future robotics and AI embodiment.
Google DeepMind has announced the launch of SIMA 2 (Scalable Instructable Multiworld Agent 2), an advanced AI agent engineered to interact with and learn within virtual 3D worlds with unprecedented human-like abilities. Unveiled on November 13, 2025, SIMA 2 represents a significant leap forward in the pursuit of Artificial General Intelligence (AGI), according to the company.
At its core, SIMA 2 is powered by Google’s sophisticated Gemini AI model, enabling it to move beyond simple instruction following. The agent can interpret complex, high-level goals, articulate its intended actions, and engage in collaborative tasks within virtual environments. This level of reasoning was not achievable by its first iteration, SIMA 1, which was released in March 2024 and primarily learned basic skills through observation and virtual keyboard/mouse controls.
DeepMind reports that SIMA 2 exhibits stronger generalization across various virtual settings and is capable of completing longer, more intricate tasks. These tasks include responding to logic prompts, sketches drawn on-screen, and even emojis. A key metric highlighting its improved performance is a 65% task completion rate, a substantial increase from SIMA 1’s 31%.
One of SIMA 2’s most remarkable features is its ability to adapt to entirely new 3D worlds. It can orient itself, comprehend objectives, and execute meaningful actions within environments generated moments earlier by Genie 3, another DeepMind project that creates interactive worlds from single images or text prompts. This demonstrates SIMA 2’s capacity to transfer learned concepts, such as applying ‘mining’ skills from one game to ‘harvesting’ in another, effectively connecting similar tasks across different contexts.
The development process for SIMA 2 involved initial learning from human demonstrations, followed by a phase of self-directed play. During this phase, the agent utilized trial and error, along with Gemini-generated feedback, to create new experience data. This included a training loop where SIMA 2 generated its own tasks, attempted them, and then fed its trajectory data back into subsequent versions of the model for continuous improvement.
Also Read:
- Google Bolsters AI Agent Safeguards with Enhanced Safety Frameworks
- Google Unveils Free 5-Day AI Agents Intensive Course on Kaggle
Despite these advancements, DeepMind acknowledges existing challenges. SIMA 2 currently struggles with very long, multi-step tasks, operates within a limited memory window, and faces visual-interpretation hurdles common to 3D AI systems. Nevertheless, the research team views SIMA 2 as a critical testbed for developing skills that could eventually be transferred to real-world applications in robotics and navigation, paving a strong path towards AGI in physical environments.
