Google reveals Gemini Robotics 2.0, promising improved dexterity and safety

Google DeepMind wants robots to stop being one-trick ponies. For years, viral videos of robots doing backflips or dancing gave the impression these machines were flexible and smart, but in reality each move was painstakingly programmed for that one specific stunt. Gemini Robotics 2 is Google's latest swing at fixing that, and it's built around an idea DeepMind researchers have started calling "physical AGI": tell a robot what you want done, in plain language, and it figures out how to do it.

The release actually bundles three new sub-models, and one of them, an upgraded embodied reasoning system called Gemini Robotics ER 2, is available to developers starting today through the Gemini Live API. That's a meaningful shift from the earlier ER 1.6 model, and DeepMind is framing it as a real leap rather than an incremental patch. The pitch is that ER 2 doesn't just execute a fixed script, it reads a changing environment continuously and adjusts on the fly, which is the kind of thing that sounds obvious for a human but has been genuinely hard for robots.

What's notable here is the scope of control. Google says this system can now manage an entire humanoid robot's body, including hands with the kind of complex joint structure that has traditionally tripped up robotic manipulation. Grasping a mug is easy for a person and brutally hard for a machine, because fingers have so many degrees of freedom to coordinate. Gemini Robotics 2 is being positioned as a step toward robots that handle that complexity without needing a bespoke program for every object they might touch.

And there's a collaborative angle too: Google is touting the ability for multiple robots running on this system to work together on a task rather than operate as isolated units. That matters for warehouses, factories, or any setting where you'd want a fleet of machines coordinating instead of colliding. None of this means general-purpose household robots are showing up next year, but the direction is clear. Google is chasing a robot that understands instructions the way a person would, not one that just replays a memorized routine.