Google DeepMind released Gemini Robotics 2 in July 2026, an AI model capable of controlling a range of robots — including humanoids that can perform dextrous tasks such as screwing in lightbulbs and tying trash bags.
The system combines multiple AI models into one. A vision language model (VLM) interprets images and video, communicates with humans, and reasons through tasks. Two vision language action (VLA) models handle physical movement, controlling both full-body motion and hand or gripper actions. Together, they allow a robot to interpret its environment and act within it.
In video demonstrations shared ahead of the release, Apptronik’s Apollo 2 robot — equipped with hands from a company called Sharpa — used the model to tidy shelves autonomously. Google DeepMind trained the system using a combination of human teleoperation, video examples, and simulations. The company notes that AI models still require specific training to perform a wide range of complex tasks.
Carolina Parada, head of robotics at Google DeepMind, described the release as “another milestone in our path towards really getting towards what we call physical AGI, which means we get a robot to do anything that a human can.” Google DeepMind CEO Demis Hassabis has previously said he hopes to develop an AI operating system for robots, similar to Android for smartphones. The company has also previously partnered with Boston Dynamics to provide AI capabilities for that company’s legged robots.
The release raises safety concerns. Prior research has shown that frontier AI controlling robots can produce unexpected and sometimes dangerous behavior. Parada acknowledged the added risk: “The safety question is even more pressing because you’re putting them in a lot of other situations.” Google says it applies safety guardrails at each model layer and is introducing ASIMOV-Agentic, a new benchmark designed to detect whether a given command could result in harmful or uncertain outcomes.
Source: WIRED