DeepMind has announced "SIMA 2," a general-purpose agent that operates in virtual 3D environments. While the previous generation of SIMA specialized in following basic linguistic instructions, the new version places Gemini models at its core. As a result, the agent has evolved from a simple command executor into an interactive game companion.
DeepMind Announces SIMA 2 Powered by Gemini, Enhancing Reasoning and Dialogue in Virtual 3D Spaces
This article is a translation. Read the Japanese original
SIMA 2 understands high-level user goals and performs complex reasoning to achieve them. Furthermore, it is capable of explaining the details of its actions and answering questions. Human demonstration videos and Gemini-generated labels were used for training.
In testing, the agent demonstrated high success rates even in unseen games. Higher reliability compared to the previous generation was observed in the Viking survival game "ASKA" and the research environment "MineDojo."
Source: SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds（HN 238pt・111コメント）（HN Search (backfill)、2025-11-14）