Google DeepMind’s new AI model can control a robot’s entire body
The Verge Emma Roth ● Covered by 7 sources
Google DeepMind's Gemini Robotics 2 can now control a humanoid robot's whole body, not just the arms. That means walking, crouching, and even tying a trash bag with five-fingered hands.
Based on reporting by The Verge, Emma Roth — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind just widened the scope of what its robotics AI can do, and the jump is bigger than it sounds. The earlier version of Gemini Robotics handled a humanoid's upper body — arms, hands, that sort of thing. Gemini Robotics 2, announced Thursday, takes over the whole machine, feet to fingertips, letting robots walk, crouch, stretch, and manipulate objects in ways the previous model simply couldn't manage.
Google shared footage of Apptronik's Apollo 2 robot bending down to grab a watering can and pulling specific items off a shelf, small tasks that nonetheless require the kind of full-body coordination that's historically been hard to pull off in humanoid robotics. The company is upfront that speed remains a weak point — these robots aren't fast — but frames the update as a real step toward machines that can handle messier, more complex real-world jobs.
Dexterity got an upgrade too. Gemini Robotics 2 can now drive more complicated five-fingered hands, which opens the door to fiddly tasks like sealing a Ziploc bag, tying off trash, or unscrewing a lightbulb — the kind of fine motor work that's easy for a human hand and stubbornly difficult for a robotic one.
Alongside the movement model, Google DeepMind updated Gemini Robotics ER, its embodied-reasoning system that helps robots read their environment and follow multi-step instructions. The new ER 2 version tracks tasks over longer stretches and, notably, now recognizes when a task actually starts and ends. It also lets different types of robots collaborate — one demo shows Apollo 2 directing a separate dual-arm robot to sort tools into a bin while tidying a garage. Google says this is its safest robotics model yet, better at spotting nearby humans and halting the robot if someone gets too close.
There's also progress on the On-Device Model, the version built to run locally without an internet connection. It now adapts faster to robots with very different shapes, sensors, and degrees of freedom, which matters if Google wants this software to work across hardware it doesn't fully control.
My take — AI-written commentary, not fact-checked reporting
Whole-body control sounds flashy, but the real tell here is Google admitting its robots are still slow — that's the honest part everyone will gloss over. Multi-robot coordination and better safety stops matter more long-term than a viral watering-can clip, and companies chasing humanoid hype would do well to notice which details DeepMind chose to caveat rather than celebrate.
Read more about this at: The Verge