Google DeepMind recently unveiled a groundbreaking artificial intelligence (AI) model, named Scalable Instructable Multiworld Agent (SIMA), which has the ability to play 3D video games with human-like proficiency. The model is currently undergoing research and training to enhance its skills in navigating diverse gaming environments and executing various tasks. Google envisions a wide range of potential applications for the SIMA model in both online and real-world scenarios.
Training and Development
DeepMind’s SIMA team emphasized that the goal is not to create a super-intelligent gaming entity that can conquer any game effortlessly. Instead, the focus is on teaching the AI model to interact in 3D open-world settings and respond to natural-language instructions similar to a human player. This presents a significant challenge as current large language models lack the capability to take physical actions based on generated plans.
To provide a comprehensive learning environment for the AI model, Google DeepMind collaborated with eight game studios and exposed SIMA to nine different video games, including popular titles like No Man’s Sky, Teardown, Goat Simulator 3, and Valheim. Each game has unique interactive elements that required the AI model to navigate, interact with objects, and use menus effectively. Additionally, the creation of research environments, such as the Construction Lab in Unity, allowed for testing the AI model’s object manipulation skills through tasks like building sculptures from blocks.
Evaluation and Skills Development
The current version of SIMA underwent evaluation across 600 basic skills, encompassing tasks like navigation (e.g., turning left, driving a car), object interaction (e.g., climbing a ladder, crafting a helmet), and more. These tasks were intentionally kept simple and short, with most activities being completed within 10 seconds. Google highlighted the potential impact of training an AI model on a diverse set of 3D games that respond to human commands, hinting at the future integration of complex strategic tasks like resource management and camp building.
As Google continues to train SIMA with progressively complex instructions demanding higher-level strategic planning and multiple sub-tasks, the company envisions practical applications that extend beyond gaming. The ability for AI models to follow human instructions and execute tasks could have significant implications in real-world scenarios, potentially aiding humans in various complex endeavors.
Google DeepMind’s introduction of the SIMA AI model represents a significant advancement in the integration of artificial intelligence in gaming and beyond. By bridging the gap between human instructions and AI actions in virtual environments, the potential for innovative applications in diverse fields is vast. As research and development progress, the future of AI-driven technologies like SIMA holds promise for enhancing human-machine interactions and problem-solving capabilities.
Leave a Reply