# Google DeepMind Adds Open Safety Benchmark to Gemini Robotics 2.5

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/54151
Published: 2026-11-11T15:31:26.000Z
Updated: 2026-11-11T15:31:26.000Z
Section: Technology

> A Mercedes-Benz proof of concept using Apptronik’s Apollo 2 recorded 96% success on pick-and-pack operations. Gemini 4 Argon scored 99.5% on the new benchmark.

Google DeepMind introduced Gemini Robotics 2.5 with a new open safety benchmark, giving robot developers a way to test whether machines obey dangerous commands disguised as harmless requests.

The release, announced Tuesday, Nov. 10, includes ASIMOV-Jailbreak, an open-source benchmark. Other labs and robot makers can use it to test their systems against unsafe instructions framed as benign requests.

Gemini Robotics 2.5 separates planning from physical execution. Gemini Embodied Reasoning takes spoken or written instructions, watches a live video feed and plans a strategy. A separate vision-language-action model turns the plan into physical movements. The update targets a wider range of tasks, faster learning, greater dexterity and improved reliability. Demonstrations included robots tying bows and using tools.

In a Mercedes-Benz proof of concept, Gemini Robotics 2.5 ran on Apptronik’s Apollo 2 humanoid robot. The system recorded a 96% success rate on pick-and-pack operations.

Gemini 4 Argon, which was used to train the robotics model, scored 99.5% on jailbreak resistance. A previous baseline scored 82%.