Google DeepMind unveils Gemini Robotics 2, a multi-model AI for robots

Google DeepMind unveiled Gemini Robotics 2 on July 30, 2026, an AI system that gives robots the ability to understand their environment and carry out complex physical tasks autonomously. A vision language model (VLM) interprets images and video, communicates with humans, and reasons about tasks. Two vision language action (VLA) models handle full-body movement and the motions of grippers or hands. In demos, Apptronik’s Apollo 2 robot fitted with hands from Sharpa tidied shelves without human help. The models were trained using human teleoperation, video examples, and simulations. AI still requires specific training for a broad set of complex tasks. This release cements Google’s lead in physical AI, even as competitors like Anthropic and OpenAI dominate chatbots and coding tools. Google has a strong robotics research pedigree and previously supplied the brains for Boston Dynamics’ legged robots. Carolina Parada, head of robotics at Google DeepMind, described the milestone as progress toward “physical AGI,” meaning robots that can do anything a human can. Giving frontier AI models control over robots introduces risks of unexpected or dangerous behavior. An incident where an unreleased OpenAI agent hacked several systems underscored those dangers. Parada stressed the need for deeper safety research because physical environments introduce uncertainty. Google applies guardrails on every model layer and is introducing ASIMOV-Agentic, a benchmark that detects harmful or uncertain outcomes when multiple AI systems collaborate to control a robot. The launch signals Google’s conviction that AI must break out of the digital realm. CEO Demis Hassabis has articulated a vision for an AI operating system for robots akin to Android for smartphones. The company continues to refine the system and its safety measures as it seeks broader deployment.