The era of robots that can only gesture from the waist up is over. Google DeepMind announced Gemini Robotics 2 on Thursday, a model that now controls a humanoid’s entire body, enabling it to walk, crouch, stretch, and manipulate objects with a full range of motion. The previous version was limited to upper-body tasks. This one goes to the mat, literally.
Videos released by the company show Apptronik’s Apollo 2 robot bending over to pick up a watering can and retrieving specific items from a shelf. The update also brings a significant leap in dexterity. The model can now control complex, five-fingered hands well enough to perform genuinely finicky tasks — think sealing a Ziploc bag, tying up a trash bag, or unscrewing a lightbulb. That’s miles away from the blunt grippers of yesteryear.
DeepMind didn’t stop at physical coordination. The companion vision-language model, Gemini Robotics ER 2, got an upgrade to its reasoning skills. It can now understand when a task begins and ends, making it better suited for long-horizon jobs. One demo shows the Apollo 2 directing a different, dual-arm robot to put tools in a bin during a garage cleanup. Multi-robot coordination is no longer just a research paper fantasy. DeepMind also flagged safety improvements, noting the model can detect nearby humans and trigger an automatic safe stop.
The company admits the robots are still a bit slow. “We have more to advance in movement speed,” the announcement notes, a rare moment of candor in a field often dominated by slick demos. There’s also an updated on-device model that can adapt faster to wildly different robot bodies without needing a cloud connection. It’s a clear signal that DeepMind is thinking about deployment, not just demos. The real question isn’t whether these robots can move — it’s whether they can do it reliably enough to leave the lab.
