DeepMind upgrades Gemini Robotics and opens safety benchmark
Gemini Robotics 2.5 expands robot dexterity while DeepMind releases ASIMOV-Jailbreak to test unsafe instructions across varied physical scenes.
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model designed to help robots acquire a broader range of physical skills and adapt faster outside controlled demonstrations. The release covers both manipulation capability and safety evaluation, including a new open benchmark called ASIMOV-Jailbreak.
From demonstrations to broader skills
DeepMind says the model improves the breadth of tasks robots can learn, the speed of skill acquisition and reliability in less structured environments. In a proof of concept with Apptronik and Mercedes-Benz, Gemini Robotics 2.5 helped an Apollo 2 humanoid robot pick and organize vehicle parts with a reported 96% success rate. Demonstrations also include fine manipulation of deformable objects, including tying a bow.
The model is intended for settings ranging from everyday assistance to factory work. That does not mean a general-purpose humanoid is ready for unsupervised deployment. Physical environments introduce failure modes that are absent from text systems: a small perception error can damage equipment, injure a person or create a cascading interruption in a production line.
Safety becomes an evaluation problem
ASIMOV-Jailbreak is an open-source benchmark containing hundreds of prompts that disguise unsafe requests across dozens of scenes. DeepMind reports that Gemini 4 Argon achieved a 99.5% jailbreak-resistance rate, compared with 82% for the previous Gemini Robotics ER 2 baseline. That result is a safety evaluation of Argon, not a reported safety score for Gemini Robotics 2.5 itself.
Publishing the benchmark gives other researchers and robot makers a common way to probe semantic safety, rather than relying only on collision avoidance or fixed motion constraints. The results remain company-reported and cover a limited set of evaluated interactions.
Why it matters
The release links robot capability to a more mature safety regime. If the benchmark gains adoption, future robotics claims may be judged not only by task success but also by how reliably a system refuses dangerous instructions in context. That is a meaningful shift for industrial deployment, where safe failure is often more valuable than a spectacular demonstration.