The researchers found that AI agents trained on multiple games performed better on games they hadn't been exposed to. However, unique mechanics or terms in many games can still confuse the AI. The researchers aim to create a more natural game-playing companion than the current hard-coded ones. They believe that SIMA players can be cooperative and can adapt and produce emergent behaviors. The researchers also noted that traditional simulator-based agent training uses reinforcement learning, which requires a reward signal for the agent to learn from. In contrast, SIMA uses imitation learning from human behavior, given goals in text.
Key takeaways:
- Google Deepmind researchers have developed a model called SIMA (scalable instructable multiworld agent) that can play multiple 3D games like a human and respond to verbal instructions.
- SIMA doesn't have access to the game's internal code or rules, but instead learns from hours of video showing human gameplay and annotations provided by data labelers.
- The researchers' goal was to see if training an AI to play one set of games makes it capable of playing others it hasn't seen, a process called generalization. The results were positive, with the AI performing better on games it hadn't been exposed to.
- The ultimate ambition is to create a more natural game-playing companion than the current hard-coded ones. The researchers envision SIMA players that are cooperative and can adapt and produce emergent behaviors.