Gemini Robotics 2.5 goes to partners, with an open robot jailbreak benchmark

Google DeepMind has introduced Gemini Robotics 2.5, which it calls its most capable vision-language-action (VLA) model so far. It is “rolling out to our robotics partners”. There is no public API, price or weights.

What DeepMind reports

• Breadth: tested on more than 200 distinct tasks, such as serving food with tongs, tying a bow on a gift box and pipetting liquid.
• Gains: 30% to 100% higher success than previous baselines depending on task type, and tasks finished 20% to 30% faster. DeepMind does not name the baseline model in that comparison.
• Factory pilot: on Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik’s Apollo 2 humanoid sorted metal brackets from bulk bins into sequencing trays with a 96% success rate. Gemini Robotics ER tracked the workflow and Gemini Robotics 2.5 drove the hands and arms.
• Robot types: Apollo 2 with a five-finger hand, and the two-gripper Franka Duo.
• Physical safety: inference-time guidance that steers around hazards raised safe task completion by 33% around obstacles and 43% around nearby people, in DeepMind’s evaluations.

DeepMind credits Gemini 4 Argon with speeding up the work by generating simulated practice runs and digital twins from real camera scenes.

The open part: ASIMOV-Jailbreak

DeepMind is releasing ASIMOV-Jailbreak, a benchmark of prompts that disguise unsafe robot requests across many scenes. The announcement links it to the google/asimov_agentic dataset on Hugging Face, which is public under CC BY 4.0 and gated behind a click-through. DeepMind says Gemini 4 Argon resists 99.5% of these jailbreaks, against 82% for its earlier Gemini Robotics ER 2. Some press reports give different figures; the ones here are from DeepMind’s post.

Why this matters

For most builders this is a direction signal, not a tool: the model is partner-only, and every performance number is self-reported. The benchmark is usable now, and if you put a language model in charge of anything with motors, disguised unsafe requests are a test you should be running. NVIDIA’s GR00T 2 is in a similar spot, with code due in December, and neither lab’s success rates have been reproduced on an open evaluation such as RoboColosseum yet.