Google Says New Gaming AI SIMA Will Act as Sandbox for More Helpful Agents

Key Takeaways

• Google Deepmind has developed a new AI Scalable Instructable Multiworld Agent (SIMA).
• SIMA learned to play video games with the goal of developing AI that can be deployed in any virtual environment.
• AI systems are increasingly built and tested with video games.

Video games are rarely accurate depictions of reality. Nonetheless, Google researchers believe they can be useful artificial intelligence (AI) training tools for real-world applications.

The company’s latest Scalable Instructable Multiworld Agent (SIMA) was designed to follow instructions and perform tasks in different video game settings.  More than just a gaming bot, SIMA is proposed as an important step toward developing a general-purpose AI agent that can operate in any virtual world.

Google SIMA Uses Video Games to Build More Helpful AI

In a research paper  published this week, the team behind SIMA emphasized that the new AI doesn’t just play games. It also understands natural language instructions and can learn new skills in diverse environments.

“A key motivation of SIMA is the idea that learning language and learning about environments are mutually reinforcing,” they said. “Even when language is not necessary for solving a task, learning language can help agents to learn generalizable representations and abstractions, or to learn more efficiently.”

Looking forward, they said the value of the study also extended beyond gaming:

“This work isn’t about achieving high game scores. Learning to play even one video game is a technical feat for an AI system, but learning to follow instructions in a variety of game settings could unlock more helpful AI agents for any environment.”

Virtual Worlds as AI Teaching Tools 

The seven games SIMA learned to play include first and third-person sandbox or open-world titles, spanning a variety of genres. 

Alongside commercial video games like Goat Simulator 3, No Man’s Sky and Valheim, the study deployed the AI in custom research environments designed to instill and assess particular skills.

Furthermore, Google isn’t the first to propose dropping AI models into virtual environments.

In February, researchers at the University of the Witwatersrand developed a new AI performance test based on Minecraft.

Known as MinePlanner, the proposed AI benchmark goes beyond existing tests used to assess language understanding and reasoning.

Observing that future AI models will need to tackle messy problems, lead researcher Steven James said researchers can use video games to test how well they deal with unfamiliar scenarios.

About the Author

James Morales

Although his background is in crypto and FinTech news, these days, James likes to roam across CCN’s editorial breadth, focusing mostly on digital technology. Having always been fascinated by the latest innovations, he uses his platform as a journalist to explore how new technologies work, why they matter and how they might shape our future.

See more

[email protected] LinkedIn Twitter