Gemini Robotics 2 controls whole robots

Google DeepMind shipped three robotics models this week that, together, move Gemini from describing what a robot should do to controlling how it does it. Gemini Robotics 2, announced July 30, is DeepMind’s first vision-language-action model that can drive an entire humanoid, not just an upper-body or tabletop arm: in DeepMind’s own demo, an Apptronik Apollo 2 robot takes a spoken instruction, walks to a table, picks up a watering can, crosses the room, and places it on a shelf, using one model checkpoint that also runs on a Franka Duo arm with a different gripper. A companion on-device variant adapts to a new robot body in a few hours of data rather than a fresh training run.

The second model, Gemini Robotics ER 2, is the reasoning layer: a vision-language model that plans multi-step tasks lasting several minutes, tracks real-time spatial state through video, and coordinates more than one robot on the same job. Developers can reach it today through the Gemini API and Google AI Studio, and it’s in a private preview inside the Gemini Enterprise Agent Platform; the action and on-device models remain limited to early-access partners for now.

What it means for operators

DeepMind is explicit that multi-finger dexterity still lags whole-body and gripper tasks, so this isn’t a drop-in replacement for teleoperation yet. But the embodiment-transfer number, adapting to a new robot in hours instead of a full retrain, is the part that changes vendor economics: it means a physical-AI integrator can stop treating every new robot chassis as a from-scratch model project. If you’re scoping a robotics pilot, ask any vendor claiming a “general” model how they handle exactly this transfer, and compare it against DeepMind’s few-hours baseline.