# DeepMind launches Gemini Robotics 2.5 and open safety benchmark

> Google DeepMind says its new Gemini Robotics 2.5 model improves dexterity, speed and safety, and a partner robot hit 96% success on Mercedes‑Benz lines.

Oossa · 2026-11-10 · https://oossa.com/en/deepmind-launches-gemini-robotics-2-5-and-open-safety-benchmark

Google DeepMind unveiled Gemini Robotics 2.5 on Nov 10, 2026. The upgrade is a vision‑language‑action (VLA) model that turns live video and text commands into robot motions. DeepMind says the new version can handle more delicate tasks, works faster, and includes safety steering that stops the robot when obstacles appear. In a pilot with Apptronik’s Apollo 2 humanoid, the system sorted automotive parts on Mercedes‑Benz production lines in Hungary and Germany and achieved a 96% success rate.

## What’s new in Gemini Robotics 2.5?

The model was tested on more than 200 distinct tasks, ranging from serving food with tongs to tying a ribbon bow and pipetting liquids. DeepMind reports a 30%‑100% higher success rate compared with the previous version and a 20%‑30% speed boost in the lab. A notable demo showed the Apollo 2 robot tying a bow with its five‑finger hand – a tricky feat because ribbons deform easily. Safety improvements come from inference‑time reasoning that guides the robot around hazards; the company measured a 33% rise in safe completions when obstacles were present and a 43% rise when people were nearby.

## Open safety benchmark

To push industry standards, DeepMind released ASIMOV‑Jailbreak, an open‑source set of 1,500 prompts designed to trick robots into unsafe actions. In internal tests the benchmark gave Gemini 2.5 a 98.8% resistance score, far above the 73.5% score of the earlier Gemini ER 2 model. The full dataset is hosted on Hugging Face for anyone working on embodied AI.

## The facts

- Gemini Robotics 2.5 was announced on November 10, 2026.
- The model was evaluated on over 200 tasks, showing 30%‑100% higher success rates than the previous version.
- In a Mercedes‑Benz pilot, Apollo 2 robots achieved a 96% success rate on automotive kitting.
- Safety steering raised safe task completion by 33% around obstacles and 43% around people.
- ASIMOV‑Jailbreak contains 1,500 safety‑testing prompts and gave Gemini 2.5 a 98.8% resistance score.

## Why it matters

For factories, the higher reliability could let robots take over repetitive, physically demanding work, freeing human technicians for more complex tasks. The open safety benchmark gives developers a way to measure and improve how robots avoid dangerous actions, which is essential before robots move out of controlled labs and into everyday spaces.

## Sources & references

1. [DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"](https://arstechnica.com/google/2026/11/deepmind-reveals-gemini-robotics-2-5-and-open-safety-benchmark-asimov-jailbreak/) – Ars Technica, 2026-11-10
2. [Gemini Robotics 2.5 accelerates real-world general dexterity](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/) – Google DeepMind, 2026-11-09

Last updated: 2026-11-10