Google DeepMind's Gemini Robotics ER 2 gives robots a sharper brain

The News

Google DeepMind on 30 July 2026 released Gemini Robotics ER 2, a high-level reasoning model that acts as a thinking layer, or brain, sitting on top of the machines that actually move. The company describes it as a significant upgrade over the earlier ER 1.6 release, with sharper video understanding, real-time spatial reasoning and the ability to coordinate several robots at once.

The headline numbers come from the model's grasp of video. On progress classification, judging how far a task has advanced across five bands from 0-20% through to 80-100%, ER 2 scores 57.4% accuracy. On moment-finding, spotting the critical instant in a clip, it reaches 91.3% accuracy with a mean absolute distance of 0.96 seconds. It can also read more than 10 kinds of instrument, from digital displays and scales to thermometers.

DeepMind says the model is available now through the Gemini API and Google AI Studio, with a private preview running on its Gemini Enterprise Agent Platform. Demonstrations were run on hardware from Boston Dynamics, whose Spot robot featured in the launch, Apptronik and its Apollo 2 humanoid, and Franka's F3 Duo collaborative arm.

Why It Matters

For two years the robotics race has been stuck on a familiar gap. Language models could describe the world fluently, but translating that fluency into safe, reliable physical action stayed stubbornly hard. ER 2 is a bet that the missing piece is not better motors but a better planner, one that watches a task unfold, notices when a step fails and decides what to do next through the Gemini Live API at sub-second latency.

The multi-robot piece is the genuinely new idea. Rather than each machine running blind, ER 2 lets robots share a semantic picture of a scene and pass work between them, one unit starting a job that another finishes. That echoes the shift software saw when single scripts gave way to orchestrated agents, and it points at warehouses and factory lines where fleets, not lone arms, do the work.

Safety was pushed to the front of the pitch. DeepMind says ER 2 beats ER 1.6 on its Safety Instruction Following and Human Proximity benchmarks, and can halt a humanoid when a person steps close. That framing matters because the last wave of humanoid hype, around 2024, foundered partly on trust rather than raw capability.

Indian Angle

India does not build frontier robots, but it is quietly building the bodies and the deployments. Bengaluru's Ati Motors and Gurugram's GreyOrange already ship autonomous mobile robots into warehouses, while Reliance-backed Addverb Technologies exports its arms and shuttles worldwide. A cheap, capable reasoning layer that plugs in over standard hardware lowers the cost of making these Indian machines smarter without building a model from scratch.

The pricing route matters just as much. Because ER 2 is exposed through the Gemini API, an Indian robotics team pays per call in the same way it would for any cloud model, sidestepping the capital cost of training. For a Genrobotics building sewer-cleaning units or a Peppermint Robotics building floor scrubbers, that turns advanced planning into an operating expense billed in rupees rather than a research programme.

There is a talent thread too. Indian engineers sit across DeepMind and the wider Alphabet stack, and domestic model efforts such as Sarvam and Krutrim are watching how far a general reasoning model can carry into the physical world. For MeitY, which is still shaping its view on AI liability, a model that can stop a humanoid near a human is a useful data point on where the safety bar should sit.

FAQ

When is Gemini Robotics ER 2 available?

It went live on 30 July 2026 through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. Developers can begin building against it immediately, subject to standard API access terms.

How is it different from ER 1.6?

ER 2 is billed as a significant upgrade, with stronger video understanding, better spatial reasoning and multi-robot coordination. DeepMind says it consistently outperforms ER 1.6 across real, simulated and teleoperated task-orchestration tests, and on its safety benchmarks.

Which robots does it run on?

Launch demonstrations used Boston Dynamics' Spot, Apptronik's Apollo 2 humanoid and Franka's F3 Duo collaborative robot. As a high-level reasoning layer it is designed to sit above varied hardware rather than a single platform.

What does it mean for Indian robotics firms?

Indian makers such as Ati Motors, GreyOrange and Addverb could add advanced planning to existing hardware via the API, paying per use rather than funding their own model training, which lowers the barrier to smarter automation.

Where can I read the original announcement?

Google DeepMind published the full details on its official blog, linked in the source attribution paragraph at the end of this article.

This story was reported by Google DeepMind. Read the full original coverage at Google DeepMind.