Google DeepMind’s Latest AI Agent Learned to Play Goat Simulator 3
Mar 14, 2024
auto_awesome
Google DeepMind's latest AI agent, SEMA, can learn to play new video games by analyzing how humans play different games. The agent has collaborated with game studios to collect data and improve its skills. DeepMind is exploring the potential for these AI agents to be used in practical tasks, while also focusing on improving their reliability.
SEMA can adapt to new games by leveraging shared concepts across different video games, showcasing remarkable skills and efficiency.
Collaboration with game studios and human evaluation feedback are crucial in training SEMA for diverse gaming environments, hinting at its future integration potential.
Deep dives
Advances in AI: SEMA's Learning Capabilities across Games
SEMA, an AI program developed by Google DeepMind, has showcased the ability to learn and adapt to complete tasks in various video games, including Goat Simulator 3. By leveraging shared concepts across games, SEMA demonstrates remarkable skills and efficiency. This project represents a significant advancement in AI, hinting at future capabilities beyond chatting and generating images, potentially enabling complex computer commands.
Training SEMA: Human-Driven Data Collection and Language Processing
The development of SEMA involved collecting human gameplay data from various 3D video games like No Man's Sky and Taredown. This data enabled SEMA to execute over 600 actions in response to human commands, ranging from exploration to tool use. Collaboration with game studios and fine-tuning based on human evaluations played a crucial role in training SEMA for diverse gaming environments, highlighting its potential for safe and reliable integration in the future.
1.
Advancements in Google DeepMind's AI Agent, SEMA, for Multi-Game Learning
These AI agents can adapt to games they haven’t played before. Google made them by feeding data on how humans play different video games to a language model like those behind the latest chatbots.