Gemini Robotics 2 Brings Whole-Body Intelligence to Humanoid Robots

Google DeepMind's Gemini Robotics 2 gives humanoids whole-body control, fine dexterity, and team coordination — a leap toward general-purpose physical AI.

Google DeepMind has unveiled Gemini Robotics 2, a major upgrade that gives humanoid robots the ability to walk, crouch, stretch, and manipulate objects with their entire body — not just their arms. For the first time, a single AI model can control a full humanoid from feet to fingertips, marking a significant step toward robots that can actually navigate the messy, cluttered spaces humans inhabit.

Three models, one intelligence layer. The release spans three complementary systems. Gemini Robotics 2 is the flagship vision-language-action model (VLA) that converts vision and language into motor commands — controlling everything from walking to five-fingered manipulation. Gemini Robotics ER 2 serves as the robot's high-level brain, handling task planning, human communication, and now multi-robot coordination. A third model, Gemini Robotics On-Device 2, runs locally on the robot itself and can adapt to entirely new robotic bodies with just a few hours of training data and fewer than 200 examples.

What it can actually do. In demonstrations, an Apptronik Apollo 2 humanoid processed the command "put the watering can into the green bin on the bottom shelf" — walking to a table, picking up the can, stepping to the shelves, and placing it precisely. A five-fingered SharpaWave hand tied knots and sealed ziplock bags. Two robots worked in tandem on tasks a single machine couldn't handle alone.

The catch. Multi-finger dexterity remains challenging. Success rates are medium-to-high for whole-body and gripper tasks, but fine finger manipulation still lags. Movement speed also needs improvement before these robots are ready for real-world deployment.

Safety built in. DeepMind is introducing ASIMOV-Agentic, a benchmark that measures whether the reasoning agent can refuse unsafe commands and know when to ask for human help. The system can detect nearby humans and trigger an automatic safe stop — a requirement for collaborative safety standards.

As commenters on Hacker News noted, this is Google competing squarely on Tesla's turf — software for physical robots rather than cars. One commenter captured the moment: "Wow, what a time to be alive." Another asked the harder question: how close is any of this to working in the wild, outside lab conditions?

The answer, for now: closer than it was. Gemini Robotics ER 2 is available on Google AI Studio today, with the VLA models accessible to early-access partners including Apptronik, Boston Dynamics, and Agile Robots. General-purpose physical AI still has a long road ahead — but it just learned to walk.

Sources: Google DeepMind blog; Hacker News discussion (455 points, 386 comments)