Archive · 14 Nov 2025
Google DeepMind Unveils SIMA 2, Gemini-Powered AI Agent for Virtual Worlds
Imagine a very smart computer player that can learn to play many different video games all by itself!
That's what Google DeepMind has created with SIMA 2.
It uses a powerful AI called Gemini to understand what you want it to do, even if it's never seen that game before.
SIMA 2 can figure out its own goals, learn from trying things over and over again, and even tell you what it's planning to do.
It's like a super helpful and smart gaming buddy that can complete many of the tasks you give it, getting better with practice.
This is a big step towards making computers that can think and learn like humans.
Google DeepMind has launched SIMA 2, an AI agent powered by Gemini, designed to operate and learn within 3D virtual environments.
SIMA 2 can follow human language instructions, reason about goals, and improve its skills through trial and error without human training data.
The agent has demonstrated near human-level performance, achieving 45-75% task completion in new, unseen games, significantly outperforming its predecessor, SIMA 1.
SIMA 2 integrates with Google's Gemini models to understand complex instructions and perform goal-oriented actions, acting as an interactive gaming companion.
This development is seen as a significant step towards Artificial General Intelligence (AGI), showcasing AI's ability to adapt to novel environments and learn independently.
- Who
- Google DeepMind
- What
- Unveiled SIMA 2, a Gemini-powered AI agent capable of following instructions, reasoning, and learning new skills in 3D virtual environments, acting as an interactive gaming companion.
- When
- Recently announced (specific date not provided)
- Where
- In 3D virtual environments and simulated games
- Why
- To advance research towards general and helpful AI agents, making significant strides towards Artificial General Intelligence (AGI).
Google DeepMind's Claims
Performance Benchmarks
Task Completion in New Games
Google DeepMind's Claims
Google claims SIMA 2 has significantly improved, nearing human-level task completion in virtual environments.
Performance Benchmarks
In unseen games like ASKA and MineDojo, SIMA 2 reportedly completed 45-75% of tasks, a substantial increase from SIMA 1's 15-30%.
Learning and Adaptation
Google DeepMind's Claims
SIMA 2 integrates advanced Gemini models, transforming from an 'instruction-follower' to an interactive companion that can reason, improve itself, and adapt to novel environments like those generated by DeepMind's Genie 3 model.
Performance Benchmarks
The AI agent was tested in generated worlds from its Genie 3 model and successfully adapted to environments it had never seen or trained on before.
Training Methodology
Google DeepMind's Claims
SIMA 2 improves through trial and error, using Gemini to create tasks and score attempts without human training data. It can also describe its intentions and steps to users.
Performance Benchmarks
The training used a mix of human demonstration videos with language labels and Gemini-generated labels.
- Product
- SIMA 2 (Scalable Instructable Multiworld Agent 2)
- Developer
- Google DeepMind
- Core Technology
- Gemini models
- Primary Function
- AI agent for learning and task completion in 3D virtual worlds
- Key Improvements
- Enhanced reasoning, interactive companionship, self-improvement, adaptation to new environments, task completion rates
- Performance
- 45-75% task completion in unseen games (compared to SIMA 1's 15-30%)
- Training Data
- Mixture of human demonstration videos with language labels and Gemini-generated labels; self-learning via trial and error
