Google: AI Agent Sima 2 Trains in Video Games for the Real World
Google has introduced a new version of its playing AI agent Sima, which masters more complex actions. Sima trains in games for the real world.
(Image: Google DeepMind)
Google's video game-playing AI agent Sima, in version 2, now also masters more complex commands and interactions with the game world. In a blog post, Google describes Sima 2 (“Scalable Instructable Multiworld Agent”) as an “interactive gaming companion” that can itself think about game objectives.
The first version of Sima was introduced by Google last year. Sima fundamentally differs from other game AIs from Google's AI division, DeepMind. While, for example, AlphaStar was specialized in defeating professional gamers in “StarCraft 2,” Sima is not about high scores in a single game. Instead, the AI model is capable of playing various games in a “natural” way.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
Accordingly, Sima's skillset can be applied to numerous different video games. According to Google DeepMind, Sima learns with every title played. A video shows how Sima 2 plays “Minecraft,” “Satisfactory,” “No Man's Sky,” and “Valheim”—all quite complex titles involving resource extraction and, in a broad sense, infrastructure building. Furthermore, Sima can interact with worlds generated by Google's model Genie 3.
Coal mining in “Minecraft”
According to Google, the first version of Sima mastered about 600 skills linked to specific instructions, such as “turn left” or “climb the ladder.” To achieve this, Sima analyzes the screen content and uses a virtual set of mouse and keyboard to interact with the game. In essence, Sima plays the way humans do. Sima 2 builds on these capabilities thanks to Gemini integration, Google writes in the blog post.
Videos by heise
The AI agent can therefore no longer just react rigidly to instructions but can also think about its commands. For example, a demo video shows Sima receiving the instruction to mine coal in a “Minecraft” cave. While Sima 1 looks for the shortest path and immediately gets stuck on a block, Sima 2 is more successful: the AI agent understands the command, navigates into the cave, and identifies coal deposits within it.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externes Video (TargetVideo GmbH) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (TargetVideo GmbH) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
Furthermore, Sima 2 can answer user queries. Upon request, the model can tell the user, for example, where the game character is currently located and why it is doing what it is doing there. However, the AI agent's goal is not to take over playing games from humans; rather, Google DeepMind sees virtual worlds as training environments where AI models can practice for the real world. Agents like Sima could, for example, be used in robots in the future to interact with real objects.
(dahe)
