Google's Gemini Robotics 2 wants to be one brain for every robot

DeepMind just shipped Gemini Robotics 2, and the pitch is blunt: one intelligence layer for any robot body. The headline upgrade is whole-body control — the previous generation mostly drove a robot's upper half, while the new model has humanoids walking, crouching and manipulating objects while reasoning through the task.

It ships as three models: the VLA for motor control, Gemini Robotics ER 2 for embodied reasoning and multi-step planning (testable today in Google AI Studio), and On-Device 2 for robots that can't phone home. DeepMind says it adapts to any bi-arm robot in a few hours — demoed on Apptronik's Apollo humanoid and the force-sensitive Franka Research 3, folding a dress and doing industrial belt assembly. Partners include Boston Dynamics, Apptronik and Agile Robots; broader access is gated behind a waitlist and 100+ trusted testers.

The honest caveat: dexterity is still the hard part, and "handles unfamiliar tasks" in a lab demo is a long way from a warehouse shift.

Why it matters: if one model really does transfer across arms, humanoids and mobile bases, robotics stops rebuilding its stack per form factor — and starts scaling like software.