Google DeepMind introduces Gemini Robotics 2.5, reporting 96% success in Mercedes-Benz humanoid pilot

Google DeepMind, Google's AI research lab, announced Gemini Robotics 2.5 on 10 November 2026, calling it its most capable vision-language-action model so far: a model that turns what a robot sees, and the instructions it is given, into movement. In a blog post by Carolina Parada, the company said the model is now rolling out to its robotics partners.

What the company reports

DeepMind says it tested the model on more than 200 distinct tasks, among them using tongs to serve food, tying a bow around a present and pipetting liquid. Against earlier baselines, the company reports success rates between 30% and 100% higher, varying by task category, and says tasks were completed 20% to 30% faster. It adds that the model runs on several kinds of robot. Two are named: Apptronik's Apollo 2 humanoid, fitted with the SharpaWave hand (five fingers, 22 degrees of freedom), and the Franka Duo, a platform that uses two-fingered parallel grippers. All of these figures are DeepMind's own.

The Mercedes-Benz proof of concept

The reliability result comes from a proof of concept with Apptronik and Mercedes-Benz on working car production lines at Kecskemét in Hungary and Sindelfingen in Germany. DeepMind says Apollo 2 carried out kitting on its own: taking single parts out of bulk bins and arranging them in trays in the order each vehicle needs. According to the company, its Gemini Robotics ER embodied reasoning model tracked the workflow while the 2.5 actions model controlled the hands and arms, so the robot could reach into supply bins, pick up metal brackets and place them in sequencing containers. DeepMind reports a 96% success rate for the pilot.

Jeff Cardenas, co-founder and chief executive of Apptronik, is quoted in the post saying the models improved sharply within weeks of post-training and carried over to the Mercedes-Benz lines with little loss in performance.

Safety measures and test set

On safety, DeepMind says Gemini Robotics 2.5 applies reasoning guidance at inference time to steer clear of physical hazards as it works. In the company's evaluations, this raised safe task completion by 33% near obstacles and by 43% near people. DeepMind also says it is releasing ASIMOV-Jailbreak, an open-source benchmark of several hundred prompts, set in dozens of scenes, that conceal unsafe requests, to test robots against adversarial jailbreaks. It reports 99.5% jailbreak resistance for its Gemini 4 Argon model, against 82% for Gemini Robotics ER 2, its earlier baseline.

The company also credits Gemini 4 Argon with speeding up development of the robotics model by serving as a simulator. It says Argon can turn one real camera scene into thousands of differing simulated runs, build a digital twin, and help researchers form hypotheses from video footage about why tasks fail.