别再肝了！Google 发布 SIMA 2，你的下一个游戏搭子可能是个 AI

Core Insights - Google has launched the next-generation general intelligence agent SIMA 2, which integrates deeply with Gemini, enabling it to understand and execute commands in virtual worlds, plan actions around objectives, and interact with players while continuously improving through trial and error [1][2] Group 1: SIMA 2 Capabilities - SIMA 2 can understand and execute complex, multi-step commands in games like "Minecraft" and "ASKA," significantly improving upon its predecessor SIMA 1, which struggled with such tasks [1][2] - The agent has been trained using a large dataset of human demonstration videos with language annotations, allowing it to develop initial "conversational collaboration" capabilities, explaining its intentions and next steps to users [2][4] - SIMA 2's task completion success rate has shown significant improvement compared to SIMA 1, demonstrating its enhanced ability to follow detailed instructions and provide feedback, akin to interacting with a real player [5][9] Group 2: Self-Improvement and Learning - SIMA 2 employs a closed-loop system of "trial and error + Gemini feedback evaluation" during training, allowing it to learn and complete more complex tasks over time [11] - The experience data accumulated by SIMA 2 can be used to train future, more powerful agents, establishing a foundation for a "general agent" capable of adapting to any world [13] Group 3: Path to General Intelligence - The combination of Gemini and SIMA 2 offers a compelling approach to achieving embodied intelligence by training agents in controlled, low-cost virtual 3D environments, where they can gather interaction data [14] - SIMA 2's ability to operate in various gaming environments is crucial for developing general embodied intelligence, enabling the agent to master skills, perform complex reasoning, and learn continuously in virtual worlds [15] Group 4: Implications for Robotics - The capabilities developed by SIMA 2, including navigation, tool use, and collaborative task execution, are essential modules for future intelligent agents to achieve "intelligent embodiment" in the real world [16]