DeepMind Launches Gemini Robotics 2.5 With 96% Success on Mercedes Lines

What happened

DeepMind said in its official blog on Nov 10 that Gemini Robotics 2.5 was evaluated across more than 200 distinct tasks covering breadth, performance and complexity, including serving food with tongs, tying a bow on a gift box and pipetting liquid; success rates ran 30% to 100% higher than earlier baselines depending on the task, with tasks completed 20% to 30% faster. The post does not name the baseline, though its charts label the comparison model as Gemini Robotics 2.

Working with Apptronik and Mercedes-Benz, Apollo 2 humanoids running the model sorted and kitted automotive parts on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, in a proof of concept with a 96% success rate. DeepMind published no trial count, success definition or rate of human intervention; Apollo 2 carries a five-finger SharpaWave hand with 22 degrees of freedom, and the model also runs on the Franka Duo dual-arm platform.

DeepMind also released ASIMOV-Jailbreak, an open-source benchmark now on Hugging Face under a CC BY 4.0 license, whose prompts disguise unsafe requests to test whether robot AI can be tricked into unsafe actions; Gemini 4 Argon resisted 99.5% of attempts, against 82% for Gemini Robotics ER 2. The model is rolling out to robotics partners only, with no public API, pricing or general availability date; Gemini 4 Argon accelerated its development through simulation.

Key facts

1. What it is: The most capable vision-language-action (VLA) model to date; camera input and instructions turn directly into arm and hand movements
2. Evaluation: More than 200 tasks; success rates 30% to 100% higher than earlier baselines depending on the task, completed 20% to 30% faster
3. Mercedes pilot: Apollo 2 kitted parts on live lines in Kecskemét and Sindelfingen at a 96% success rate; no trial count or human-intervention rate published
4. Hardware: Apollo 2's five-finger SharpaWave hand, with 22 degrees of freedom; also runs on the Franka Duo dual-arm platform
5. Safety benchmark: ASIMOV-Jailbreak is open source on Hugging Face under CC BY 4.0; Gemini 4 Argon resisted 99.5% of jailbreak attempts, Gemini Robotics ER 2 resisted 82%
6. Availability: Rolling out to robotics partners only; no public API, pricing or general availability date