Google DeepMind has released Gemini Robotics 2.0, featuring three new sub-models designed to create more capable generalist robots. One sub-model is publicly available for developers starting today. The release aims toward "physical AGI"—robots that can perform any task a human could do, rather than being limited to narrow pre-programmed actions.
The new models offer improved dexterity for humanoid robots with complex hands and better video analysis capabilities. The embodied reasoning model (ER 2) can identify key moments in video feeds with approximately 90 percent accuracy, such as knowing when to stop pouring coffee. It also allows robots to understand failures in real time and retry individual steps rather than restarting entire tasks.
Gemini Robotics 2.0 enables robot collaboration, with demos showing Apptronik's Apollo 2 and Franka F3 Duo working together without interference. The action models are described as more accurate and efficient, with a smaller on-device version adaptable to new robot designs using roughly 200 movement examples. Google has also introduced a new safety benchmark called ASIMOV-Agentic to evaluate models across various safety factors, reflecting the company's stated commitment to combining traditional physical safety measures with robust AI safety frameworks.