Archive · 15 Nov 2025
Google DeepMind Unveils SIMA 2, an Advanced AI for Game Exploration and Task Completion
Imagine a smart computer helper that can play video games!
Google DeepMind made a new version called SIMA 2.
It's like a robot brain that can understand what you tell it to do, like 'find the red house' or 'build a fort'.
SIMA 2 looks at the game world on the screen and figures out all the little steps needed to finish the job.
It's really good at learning new games it's never seen before and can even use what it learned from one game to help it in another.
It can understand instructions given in different ways, like words or even drawings.
But, it's not perfect yet.
Sometimes it forgets things it did before, and it has trouble planning really long tasks with many steps.
Google DeepMind has launched SIMA 2, an upgraded AI agent for gaming, enhancing reasoning and adaptability.
SIMA 2 is powered by Google's Gemini models and learns by following human instructions and observing its virtual environment.
The AI can break down complex goals into smaller actions and execute them using keyboard and mouse-like controls.
SIMA 2 shows improved ability to perform in new, unfamiliar games and can transfer knowledge between different game types.
Limitations include restricted memory of past interactions and challenges with long-range planning and precise low-level control.
- Who
- Google DeepMind
- What
- Introduced SIMA 2, an advanced AI agent designed to understand instructions, plan, and explore within video games like a human.
- When
- Announced on Thursday.
- Where
- Tested in game environments such as Minedojo (Minecraft adaptation) and ASKA.
- Why
- To advance AI capabilities in understanding and interacting with complex virtual environments, learning from instructions and demonstrated actions.
AI Capabilities
AI Limitations
Adaptability and Learning
AI Capabilities
SIMA 2 demonstrates significant improvements in adaptability, learning continuously and becoming more capable through its own play. It can operate in games it has not encountered before, applying concepts learned in one game to another.
AI Limitations
SIMA 2 has notable limitations, including restricted memory of past interactions and difficulty with long-range reasoning that requires many steps. Precise low-level control is also not addressed.
Task Understanding and Execution
AI Capabilities
Powered by Gemini models, SIMA 2 can reflect on its actions, think through steps to complete tasks, understand human-issued instructions, and plan moves based on visual input. It breaks down complex goals into smaller actions.
AI Limitations
The current framework does not address precise low-level control, which would be similar to robotic joint movements, indicating a gap in fine-grained physical manipulation within virtual environments.
Interaction and Input
AI Capabilities
SIMA 2 handles multimodal prompts including sketches, emojis, and various languages, enhancing user interaction.
AI Limitations
Memory of past interactions is restricted, limiting its ability to recall and build upon previous experiences or dialogues.
- AI Agent Name
- SIMA 2 (Scalable Instructable Multiworld Agent)
- Developer
- Google DeepMind
- Launch Date
- Thursday (implied April 11, 2024, based on article date)
- Underlying Technology
- Google's Gemini models
- Key Improvement
- Enhanced reasoning, adaptability, and user interaction, with improved performance in novel game environments.
- Training Method
- Mix of human demonstration data and automatically generated annotations from Gemini models.