Sciencepulse

DeepMind Unveils Gemini Robotics 2.5 and an Open Robot Safety Benchmark

Robotics·November 11, 2026

DeepMind Unveils Gemini Robotics 2.5 and an Open Robot Safety Benchmark

Google DeepMind has announced Gemini Robotics 2.5, the latest version of its model for controlling physical robots. The company says the model is up to twice as successful on some tasks. The announcement, as summarized, does not specify which tasks were tested or what the comparison baseline was, so those claims should be read as the company's own figures until independent results appear.

Alongside the model, DeepMind released ASIMOV-jailbreak, an open benchmark focused on safety. The name points back to the Three Laws of Robotics that Isaac Asimov wrote about in his science fiction, and DeepMind's earlier ASIMOV work already tested whether robots respond safely to instructions. Going by its name, a jailbreak benchmark checks whether a model can be talked into ignoring its safety rules, which is a practical concern once language-driven models are connected to machines that can move objects.

The release reflects a broader shift. As AI systems move from chat windows into hardware, a failure can mean damage to property or harm to people, not just a wrong answer. Making the safety test open means outside researchers can run it themselves rather than relying only on the lab's own reporting. Anyone following the story should watch for independent evaluations of both the performance gains and the safety results.

Reporting based on an external source.