# Gemini Robotics 2.5: what Google's robot AI can do

Google DeepMind's Gemini Robotics 2.5 controls robot arms and hands. What's new, the Mercedes-Benz test, who can use it and what it means for you.

URL: https://aitoolsthisweek.com/gemini-robotics-2-5
Published: 2026-11-15T03:27:04.795Z
Updated: 2026-11-15T03:27:04.795Z

**Gemini Robotics 2.5 is Google DeepMind's most capable AI model for robots yet: it looks at a scene, takes an instruction, and works out how a robot's arms and hands should move to do the job.** It was announced on [November 10, 2026](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/) and is rolling out to robotics partners, not to the public.

It's trending on Product Hunt this week, so here's what's actually new and what it means if you don't build robots.

## What Gemini Robotics 2.5 is

DeepMind calls it a *vision-language-action* (VLA) model. In plain terms:

- **Vision:** it sees through the robot's cameras
- **Language:** it understands a spoken or written instruction
- **Action:** it outputs the movements for arms, hands and grippers

It works as a pair. **Gemini Robotics ER**, a reasoning model, plans the job and keeps track of progress, while Gemini Robotics 2.5 handles the fine motor control. DeepMind also used **Gemini 4 Argon** as a kind of simulator: it turned real camera scenes into practice runs and digital twins and helped find why tasks failed.

## What's better than before

DeepMind's reported numbers, compared with its earlier baselines:

| Measure | Result |
|---|---|
| Tasks tested | More than 200 |
| Success rate | 30% to 100% higher, depending on the task |
| Speed | Tasks finished 20% to 30% faster |
| Safe completion around obstacles | Up 33% |
| Safe completion around nearby people | Up 43% |

These are DeepMind's own tests, not independent benchmarks.

## The Mercedes-Benz factory test

With robot maker Apptronik, DeepMind ran Apollo 2 humanoid robots on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The job was *kitting*: picking car parts out of bins and placing them in order on trays. DeepMind reports a **96% success rate**. The robots used a five-finger hand with 22 degrees of freedom. The model also ran on the Franka Duo platform, which uses simple two-finger grippers.

## Safety

Alongside the model, DeepMind released **ASIMOV-Jailbreak**, an open benchmark that tests whether a robot AI can be tricked by prompts that disguise unsafe requests. It reports 99.5% resistance for Gemini 4 Argon, up from 82% for Gemini Robotics ER 2.

## What it means for you

You can't download it or try it in an app. The near-term impact is in factories and warehouses, where robots that learn new tasks faster could take on repetitive work like sorting parts. Home robots that fold laundry or tidy a kitchen are still not something Google is selling, and the announcement gives no consumer timeline.

To follow the models behind it, read our explainers on [Gemini 4 Hydrogen](/gemini-4-hydrogen) and [Gemini 4 Carbon](/gemini-4-carbon). New to the idea of AI that acts on its own? Start with [what is an AI agent](/what-is-an-ai-agent).

*AI Tools This Week is not affiliated with Google.*

## Frequently asked questions

### Can I use Gemini Robotics 2.5?

Not as a consumer. Google DeepMind says it's rolling out to robotics partners, and the announcement doesn't mention a public API, Google AI Studio access or a price.

### Is Gemini Robotics 2.5 the same as the Gemini chatbot?

No. It belongs to the Gemini family, but it's a vision-language-action model built to move robot arms and hands, not to chat. A separate reasoning model, Gemini Robotics ER, plans the task and tracks progress.

### Is Gemini Robotics 2.5 safe around people?

DeepMind reports a 43% improvement in completing tasks safely around nearby people and 33% around obstacles, and it released an open safety test called ASIMOV-Jailbreak. These are DeepMind's own results, not independent ones.

## Sources

- [Gemini Robotics 2.5 accelerates real-world physical intelligence (Google DeepMind, Nov 10, 2026)](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)
