Gemini Robotics 2 brings whole body intelligence to robots
Google DeepMind ● Covered by 7 sources
Google DeepMind launched Gemini Robotics 2, a set of AI models giving robots whole-body control, dexterous hands, and teamwork skills. It's a big step toward robots that adapt on the fly instead of just running pre-set routines.
Based on reporting by Google DeepMind — 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 pushed its robotics AI a lot further than the tabletop tricks we've seen before. The new release, Gemini Robotics 2, is actually three models working together: a vision-language-action model that controls movement, an embodied reasoning model that plans and talks to humans, and an on-device version built for robots without reliable internet.
What stands out is the shift from arms to whole bodies. Earlier versions of Gemini Robotics handled upper-body tasks on a tabletop. Now, paired with Apptronik's Apollo 2 humanoid, the system can take a spoken instruction like putting a watering can into a green bin on a bottom shelf, and actually walk to the table, pick the can up, cross the room, and place it correctly. Google admits movement speed still needs work, but coordinating legs, arms, and decision-making in one instruction is a genuine leap from scripted, repetitive robot behavior.
Dexterity gets a boost too. The model can drive Apollo 2's five-fingered SharpaWave hand, which has 22 degrees of freedom, through fiddly jobs like tying knots or sealing a ziplock bag. It also works with simpler two-fingered grippers on a Franka Duo setup for tasks like tight packing, so the same intelligence layer scales across very different hardware.
The reasoning model, Gemini Robotics ER 2, is meant to act as the robot's brain for longer jobs, tracking multi-step tasks that run several minutes and involve hundreds of decisions, and correcting course when something goes wrong. Google is also introducing multi-robot collaboration, letting separate robots split up a workflow that one machine couldn't finish alone. Meanwhile, the on-device model can adapt to new bi-arm robot bodies in just a few hours using fewer than 200 examples, even when the new hardware has a completely different shape or sensor setup than what it trained on.
Safety gets real attention here rather than an afterthought. Google introduced a benchmark called ASIMOV-Agentic to test whether the reasoning model refuses unsafe commands and knows when to ask a human for help instead of guessing. The company says Gemini Robotics ER 2 is its safest robotics model yet on proximity and safety-constraint tests, including stopping a robot when a person gets too close.
My take — AI-written commentary, not fact-checked reporting
Robots that can walk across a room and finish an actual task, instead of just posing for a demo video, is the part that matters here. The dexterity and whole-body control are impressive, but the real test is whether this generalizes outside Google's curated setups with Apptronik and Franka hardware, because plenty of robotics demos look magical in a lab and fall apart the moment the shelf is a different height. The safety benchmark is a smart addition too — nobody wants a humanoid that only learns caution after the recall notice.
Read more about this at: Google DeepMind