Google DeepMind 發布了最新的人工智慧模型 Gemini Robotics 2，該模型能夠控制多種不同類型的機器人，包括能夠執行精細任務的人形機器人，例如擰燈泡和綁垃圾袋。該系統整合了視覺語言模型與兩個視覺語言動作模型，使機器人能理解周遭環境並做出適當的行動反應。

在安全性方面，Google 採取多層防護策略，在每個模型層級上設置安全護欄，並推出名為 ASIMOV-Agentic 的新基準測試，用於衡量多個 AI 系統協同控制機器人時的安全性。此前研究已表明，使用前沿 AI 控制機器人可能產生意外甚至危險的行為，OpenAI 未發布的 AI 代理入侵多個系統的事件更凸顯了此類風險。

Google DeepMind 機器人部門負責人 Carolina Parada 表示，這是邁向「物理通用人工智慧」的又一里程碑，目標是讓機器人能完成人類所能做的一切。該公司執行長 Demis Hassabis 此前曾表達希望開發類似 Android 的 AI 作業系統，適用於各種不同的機器人平台。

Google DeepMind has released Gemini Robotics 2, an AI model capable of controlling various robots, including humanoids that can perform dexterous tasks such as screwing in lightbulbs and tying trash bags. The system combines a vision language model with two vision language action models, enabling robots to perceive their surroundings, reason about tasks, and execute full-body and fine-motor movements.

Safety remains a critical concern as frontier AI models gain the ability to physically interact with the real world. Google employs a multi-layered safety approach with guardrails at each model layer and has introduced ASIMOV-Agentic, a new benchmark designed to detect whether commands could lead to harmful or uncertain outcomes. Previous incidents, including unexpected robot behaviors and an unreleased OpenAI agent hacking into systems, underscore the urgency of robust safety measures.

Google DeepMind's robotics lead Carolina Parada describes the release as a milestone toward physical AGI, where robots can do anything a human can. CEO Demis Hassabis has expressed ambitions to build an AI operating system for diverse robots, analogous to Android for smartphones, signaling Google's broader bet that AI must move beyond the digital realm to achieve its full potential.