The significance of Google DeepMind's latest robotics release lies less in any single feat than in the direction of travel.
DeepMind, the artificial intelligence lab owned by Google, wants to build for robots what large language models became for text: a general-purpose base that many machines can share.
Gemini Robotics 2 is a set of models that can control an entire humanoid body, not just a table-top arm.
The system coordinates movement across limbs, handling walking, bending and five-fingered grasping in what the company calls whole-body motion from feet to fingertips.
DeepMind demonstrated the models on Apollo, a humanoid built by the US robotics firm Apptronik, bending, walking and picking items from a shelf.
One brain, many bodies
The deeper change is what DeepMind calls transfer across embodiments, meaning the same underlying model can run on different robot designs.
That matters because improvements can spread across machines rather than being rebuilt for each one, the pattern that made foundation models so powerful in software.
Humanoids have long been held back less by motors and joints than by the software needed to make sense of a messy physical world.
By treating that software as a shared, improvable model, DeepMind is betting the economics that transformed chatbots can be applied to physical labour.
The release arrives as a family of three: a motor-control model, an embodied-reasoning model that plans multi-step tasks, and an on-device version that runs locally on the robot.
Running on the robot itself cuts the delay of sending data to the cloud and keeps the machine working without a network connection, useful in factories and homes.
The reasoning model, Gemini Robotics ER 2, adds longer task planning and safeguards that better detect nearby people and can trigger a safe stop.
DeepMind bills it as its "safest robotics model to date", a claim that matters because working near humans is the main barrier to deploying robots commercially.
And the rest?
The update sharpens a contest that already includes Tesla, the electric carmaker building its Optimus humanoid, alongside startups such as Figure AI and the established maker Boston Dynamics.
DeepMind's edge is software, since it supplies the intelligence while partners such as Apptronik supply the hardware.
The commercial prize is large, with humanoids pitched for warehouses, factories and eventually homes where labour is scarce or repetitive.
For Google, the move extends Gemini from a chatbot brand into a platform that could underpin a coming wave of machines.
The company is candid that the work is unfinished, admitting "we have more to advance in movement speed" before the models are production-ready.
Today's robots remain slower and less agile than the demonstration videos imply, and multi-robot collaboration is still an ambition rather than a shipped product.
For now the relevance lies in the trajectory, a push towards general-purpose robot intelligence that transfers across machines, even as the bodies it controls struggle to keep up.
