Gemini Robotics 2 gives humanoids whole-body control — demonstrated on Apptronik's Apollo 2

Google DeepMind announced Gemini Robotics 2, a set of physical-AI models that extend robot control from the upper-body manipulation of earlier versions to whole-body coordination — what DeepMind describes as "feet to fingertips." The flagship demonstration runs on Apptronik's Apollo 2 humanoid: told to "put the watering can into the green bin on the bottom shelf," the robot walks across the room, locates the object, crouches and places it, reasoning through the steps in real time. Dexterity tests using five-fingered hands from Sharpa covered knot-tying, sealing zip-lock bags, loading tapes into a boombox and unscrewing a lightbulb at a 92% success rate. A companion model, Gemini Robotics ER 2, acts as an embodied reasoning brain for chat, multi-step planning, progress tracking, tool use and multi-robot collaboration.

Why it matters

Humanoid hardware has outrun humanoid software for two years — the bottleneck has been control, not actuators. Whole-body coordination is the specific unlock that turns a robot from a fixed-position manipulator into something that can cross a room and finish a task, which is the difference between the factory demos this newsroom has logged and general deployment. For Apptronik it converts a robot-park buildout into a credibility claim against Figure at BMW and Tesla's Optimus; for the market it arrives days after Washington walled off Chinese humanoids, meaning US makers now have both a protected market and a frontier control stack — an answer to the argument that Chinese hardware advantage was decisive. The caveat is the usual one: these are vendor demonstrations with selected success rates, not deployment statistics.

What to watch

Whether third parties reproduce the dexterity numbers outside DeepMind's demos, which humanoid makers beyond Apptronik get access to the models, and the first whole-body deployments doing paid work rather than showcase tasks.