Gemini Robotics 2.5 reaches 96% success on Mercedes-Benz lines

By Grace Kim

Google DeepMind's Gemini Robotics 2.5 reached a 96% success rate in a Mercedes-Benz factory proof of concept. The model turns camera images and instructions into arm and hand movements.

DeepMind reported success rates 30% to 100% higher across more than 200 tasks, depending on the task type. Tasks also ran 20% to 30% faster than earlier baselines.

What changed in Gemini Robotics 2.5?

Google DeepMind released the model on Tuesday. The company called it its most capable vision-language-action, or VLA, model to date. Access begins with DeepMind's robotics partners. There is no public API. The system converts camera images and instructions into arm and hand movements.

DeepMind's tests included serving food with tongs, tying a bow on a gift box and pipetting liquid. The company reported the largest gains on deformable objects and multi-step sequences that July's Gemini Robotics 2 release mostly could not complete.

Task success rates from DeepMind's charts:

• Pack pastries with tongs: 5% for Gemini Robotics 2; 100% for Gemini Robotics 2.5.
• Close a toolbox lid: 0% for Gemini Robotics 2; 93% for Gemini Robotics 2.5.
• Untie ribbon bows: 0% for Gemini Robotics 2; 70% for Gemini Robotics 2.5.
• Hang a cardigan: 0% for Gemini Robotics 2; 62% for Gemini Robotics 2.5.
• Sweep trash into a dustpan: 0% for Gemini Robotics 2; 52% for Gemini Robotics 2.5.
• Fold pants: 0% for Gemini Robotics 2; 40% for Gemini Robotics 2.5.

What did the Mercedes-Benz pilot show?

DeepMind, Apptronik and Mercedes-Benz ran Apollo 2 humanoids on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, as part of a proof of concept.

Gemini Robotics ER tracked the workflow. Gemini Robotics 2.5 controlled the hands and arms. The robots pulled metal brackets from bulk bins and placed them in sequencing containers.

DeepMind described the factory work as a proof of concept.

Kanishka Rao, a DeepMind director, told Ars Technica that closing the final 4% would be a "steep road." The model still runs on cloud servers, which adds latency.

Jeff Cardenas, Apptronik's CEO, said the models "improve dramatically in just a few weeks of post-training" on Apollo 2.

How does Google DeepMind test robot jailbreaks?

DeepMind used Gemini 4 Argon as a simulation engine during development. It generated thousands of varied practice runs and digital twins from real camera scenes. The system also helped diagnose failures from video.

Safety guidance applies reasoning at inference time to steer around hazards. DeepMind reported a 33% increase in safe task completion around obstacles. Safe completion around nearby people rose 43%.

DeepMind's open ASIMOV-Jailbreak benchmark tests whether robot AI can be tricked into unsafe actions through prompts that disguise unsafe requests across dozens of scenes.

It is available on Hugging Face under a CC BY 4.0 license. Gemini 4 Argon resisted 99.5% of those attempts, against 82% for Gemini Robotics ER 2. The results leave a 17.5-percentage-point difference in resistance to those disguised prompts.

What happens next?

The release follows Nvidia's GR00T 2 robot model, which learns new skills from human video.

Gemini Robotics 2.5 reaches DeepMind's robotics partners, with no public API. Rao identified the final 4% and cloud latency as unresolved issues.

The next test for partners is whether they can close the final 4% while limiting cloud latency and preserving the reported task speed and safety gains.