DeepMind unveils Gemini Robotics 2.5, with humanoid hitting 96% success in Mercedes-Benz pilot

Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, a vision-language-action model for controlling robots. The company showed it in a pilot on active Mercedes-Benz production lines, where Apptronik's Apollo 2 humanoid sorted parts with a 96% success rate, according to the company.

What was announced

According to a DeepMind post by Carolina Parada, Gemini Robotics 2.5 is the lab's most capable VLA model to date. Development focused on three areas: dexterity, reliability and safety. The model works across different robots, including Apptronik's Apollo 2 humanoid with a five-finger hand and the Franka Duo platform with two-fingered grippers.

DeepMind says the tests covered more than 200 tasks, including serving food with tongs, tying a bow on a gift box and pipetting liquids. They required tool use and handling deformable and delicate objects.

The test numbers

The company claims a 30-100% higher success rate and 20-30% faster completion than earlier baselines, depending on the type of task.

These are the manufacturer's figures and have not been independently verified.

Pilot in Mercedes factories

Apollo 2 was tested on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task is kitting: the robot picks parts from bulk bins and arranges them in trays in the order needed for each vehicle. The Gemini Robotics ER model tracked the workflow, while the 2.5 actions model controlled the hands and arms, placing metal brackets into sequencing containers.

The success rate was 96%, according to Google DeepMind; The Decoder reports the figure as reported by the company. DeepMind describes the pilot as promising progress toward widespread industrial applications.

Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off. - Jeff Cardenas, co-founder and CEO, Apptronik

Safety and ASIMOV-Jailbreak

Inference-time reasoning is meant to steer the robot around hazards: DeepMind reports a 33% improvement in safe task completion around obstacles and 43% around nearby people. It also released ASIMOV-Jailbreak, an open benchmark with hundreds of prompts that disguise unsafe requests across dozens of scenes.

The 99.5% jailbreak resistance figure applies to Gemini 4 Argon, not Gemini Robotics 2.5; the earlier Gemini Robotics ER 2 scored 82%. Argon also served as a simulation engine: it generates thousands of practice-run variants from a single camera scene and builds a digital twin, which is meant to help robots generalize to new scenes.

What comes next

The model is rolling out to Google's robotics partners, including Agile, Apptronik and Mercedes-Benz. DeepMind calls the release a milestone on the path to solving "physical AGI", which is the company's own assessment. For Polish industrial firms, this is for now a signal from European factories, not a product available on the market.