Gemini Robotics 2 Gives Humanoids Whole-Body Control

Picture the operations manager of a mid-sized warehouse watching a demo video this week: a humanoid robot picks up a watering can, walks it across the room, and slots it neatly into a shelf bin. No teleoperator, no scripted path. The video came from Google DeepMind, which unveiled Gemini Robotics 2 on July 30, and it is the closest thing yet to a robot that simply does what you ask.

The release is a stack of three models. At the center sits a vision-language-action model that converts what the robot sees and hears directly into motor control, steering a full humanoid, in DeepMind's phrase, "from feet to fingertips," along with various bi-arm platforms. Demonstrations ran on Apptronik's Apollo 2 humanoid, the Franka Duo, and bi-arm systems from Dexmate, Trossen, and others. The third member of the stack is an On-Device variant that runs on the robot's own hardware without a cloud connection, aimed at settings where latency or connectivity cannot be trusted.

The planning brain is the quiet headline

Gemini Robotics ER 2 handles the reasoning layer: it interprets instructions, observes the environment, breaks a job into steps, and coordinates with the action model. New this generation is multi-robot collaboration, with different machines communicating and splitting a task between them. Fleet orchestration, rather than single-robot dexterity, may turn out to be the feature that matters commercially, because real facilities run many machines on one job, and the coordination has always been the hard part.

The competitive angle is worth a sentence too: Figure, Tesla, and a crowd of Chinese manufacturers all build their own control software for their own machines. DeepMind builds no robot at all; it is selling the intelligence layer across everyone else's hardware, and if that bet lands, it ends up in an Android-like position over the whole category.

Access is staged. ER 2 is available in Google AI Studio and in private preview on the Gemini Enterprise platform; the action models remain limited to early-access partners, with a Trusted Tester sign-up for everyone else.

A buyer's-eye view

Back to that warehouse manager: there is nothing to purchase this quarter, and anyone promising otherwise is selling ahead of the product. The trend underneath is worth planning around, though. Robot intelligence is detaching from robot hardware and moving into APIs, which means the expensive part of a future deployment will be integration rather than mechanics. Businesses that get their processes and data pipelines API-ready today will be the ones that move quickly when the hardware catches up, and the ones watching demos on repeat while competitors deploy.

Sources: Google DeepMind, Robotics & Automation News