Google DeepMind has introduced SIMA 2, an evolution of its previous agent, designed to navigate and solve tasks in complex virtual environments. Leveraging the Gemini large language model, SIMA 2 enhances its capability to interact, learn from experience, and take user instructions through chat or drawing. Unlike previous game-playing agents, this system aims to operate in an open-ended gaming environment where it forms its own goals based on user input.
Researchers claim that SIMA 2 significantly improves its task performance, learning through trial and error, and adapting to new environments generated by Genie 3, a world model. Despite these advancements, SIMA 2 faces challenges with complex multi-step tasks, and its memory limitations restrict its ability to retain information over time. Skepticism exists around the transferability of skills learned in gaming to real-world applications, raising questions about the overall effectiveness of SIMA 2 as a foundation for future robotics. Nevertheless, Google DeepMind remains optimistic about the agent’s potential to evolve through continuous testing and feedback in virtual training environments.
👉 Pročitaj original: MIT Technology Review – AI
