Google DeepMind's Gemini Robotics 2 hands humanoids full-body control

The News

Google DeepMind has unveiled Gemini Robotics 2, an upgraded family of models that lets humanoid machines coordinate their entire bodies, from feet to fingertips. Announced on Thursday, the release marks a jump from the firm's earlier work, which concentrated on a robot's upper body, to systems that can bend, reach, balance and walk while carrying out a task.

The launch comes in two parts. Gemini Robotics 2 is a vision-language-action model that turns what a robot sees and hears into motor control, letting it handle jobs it was never explicitly trained on. Its companion, Gemini Robotics ER 2, is an embodied-reasoning model that plans within physical space, breaks long jobs into steps and can even divide labour across several robots at once.

The dexterity on show is fine-grained. DeepMind says the models can manage precision work such as screwing in a bulb or tying a knot. In demonstrations, hardware from research partners Agile Robots, Apptronik and Boston Dynamics ran the software, with Apptronik's Apollo unit lifting objects from a shelf. More than 100 organisations are now testing the models through a trusted-tester programme, and Gemini Robotics ER 2 can be tried inside Google AI Studio, with a waitlist open for wider access.

Why It Matters

Whole-body control is the line that separates a laboratory demonstrator from a machine that can earn its keep on a factory floor. When DeepMind first brought Gemini to robots in early 2025, the models were impressive but bench-bound, good at manipulating objects placed within easy reach. Teaching a robot to move its legs, shift its weight and stay upright while its hands work is a far harder problem, and it is the one standing between today's humanoids and real deployment.

The timing sharpens a contest that already involves Tesla's Optimus programme, Figure and a crowd of well-funded startups. The pattern echoes the large-language-model race of 2023, when a capable general model turned a research curiosity into a product category almost overnight. A shared robotics brain that any partner can license could do the same for humanoid hardware, shifting the value from the metal to the software that drives it.

Indian Angle

For India, the news lands on a sensitive fault line. The country's manufacturing pitch, from the production-linked incentive schemes to the courting of firms like Foxconn and Tata Electronics, rests heavily on a large, comparatively low-cost workforce. Capable whole-body humanoids threaten to erode that edge over time, while also offering a route for Indian factories to scale output without importing labour-market bottlenecks.

The domestic robotics scene is not starting from zero. Noida-based Addverb Technologies, backed by Reliance, and GreyOrange have built warehouse-automation businesses, while Bengaluru's CynLr is working on the exact problem DeepMind is chasing, giving machines the visual dexterity to pick up unfamiliar objects. A licensable foundation model raises a strategic question for them: build the brain, or build the body and buy the brain from Mountain View.

There is also a policy gap. MeitY has moved on AI governance and data rules, but India has no dedicated framework for autonomous physical robots in workplaces, and no clear labour-transition plan should adoption accelerate. For Indian developers, access via Google AI Studio also carries a rupee cost that will shape who gets to experiment early.

FAQ

When was Gemini Robotics 2 announced?

Google DeepMind revealed the models on Thursday, positioning them as its latest robotics breakthrough alongside the embodied-reasoning model Gemini Robotics ER 2.

How is it different from the earlier Gemini Robotics model?

The first release focused largely on a robot's upper body and hands. Gemini Robotics 2 adds whole-body control, coordinating movement from feet to fingertips so a humanoid can balance, bend and walk while working.

Which robots can run it?

DeepMind demonstrated the models on hardware from Agile Robots, Apptronik and Boston Dynamics, including Apptronik's Apollo humanoid.

Can Indian developers access it?

Gemini Robotics ER 2 is available to try in Google AI Studio, and a waitlist is open for the trusted-tester programme, which already spans more than 100 organisations worldwide.

Where can I read the original announcement?

The full coverage is linked in the source attribution below.

This story was reported by The Verge. Read the full original coverage at The Verge.