Gemini Robotics 2 brings whole body intelligence to robots

Google DeepMind dropped Gemini Robotics 2 today, and the headline shift is that robots controlled by it can now move their entire bodies, not just wave their arms around a tabletop. Earlier versions of Gemini Robotics handled upper-body manipulation fine, but anything below the waist was someone else's problem. Now the model can tell a humanoid like Apptronik's Apollo 2 to walk across a room, pick up a watering can, take a few more steps, and set it down precisely on a shelf. That's a small errand for a human. For a robot's control stack, it's a genuinely hard coordination problem involving balance, navigation, and grasping all at once.

The dexterity upgrade is arguably the more interesting part. Gemini Robotics 2 can drive Apollo's five-fingered SharpaWave hand, which has 22 degrees of freedom, well enough to tie knots or seal a ziplock bag. It also works with plain two-finger grippers on a Franka Duo arm for tight-packing tasks. That range matters because most real robots in warehouses and factories still use simple grippers, not humanoid hands, so a model that generalizes across both hardware types is more useful than one built for a single showcase robot.

DeepMind split the system into three pieces. Gemini Robotics 2 itself is the vision-language-action model that turns instructions into motor commands. Gemini Robotics ER 2 is the reasoning layer, acting like a project manager that breaks a multi-minute task into steps, tracks progress, and now lets multiple robots coordinate on a job none of them could finish alone. Gemini Robotics On-Device 2 is the lightweight version meant for robots without reliable network connections, and it can apparently adapt to a brand-new robot body — different arms, different sensors — using under 200 examples and just a few hours of tuning. That fast-adaptation claim, inherited from the 1.5 generation's