Products and deployments / Manufacturing
Gemini Robotics 2.5 goes to robot partners, with a 96% success rate in a Mercedes-Benz parts-kitting pilot
Google DeepMind said Nov 10, 2026 that its new robot control model is rolling out to partners. The 96% figure comes from a proof-of-concept with Apptronik's Apollo 2, and the post gives no trial count.
Google DeepMind published a post on Nov 10, 2026 introducing Gemini Robotics 2.5, which it calls its most capable vision-language-action model to date. It says the model "is rolling out to our robotics partners." It links "our robotics partners" to DeepMind's Gemini Robotics page, which lists Agile Robots, Apptronik and Boston Dynamics as research partners. The post gives no general release date.
The post says DeepMind partnered with Apptronik and Mercedes-Benz to run Apollo 2 humanoid robots on active automotive production lines in Kecskemét, Hungary, and Sindelfingen, Germany. In the proof-of-concept, Apollo 2 does automotive kitting: it picks individual metal brackets from supply bins and places them in sequencing containers in the order each vehicle needs. DeepMind says the pilot reached a 96% success rate.
The post does not say how many picks or runs the 96% covers, how long the pilot ran, how success was scored, or whether people stepped in. It does not say whether the robots remain in use at either plant. It carries no quote from Mercedes-Benz, which it thanks in its acknowledgements. Apptronik CEO Jeff Cardenas is quoted saying the models "improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off."
Outside the factory pilot, DeepMind says it tested the model across over 200 distinct tasks, such as serving food with tongs and pipetting a liquid. Against previous baselines it reports a 30% to 100% higher success rate depending on task type, and tasks finished 20% to 30% faster. It names Apptronik's Apollo 2 and the Franka Duo platform among the robot types it works with.
On safety, the post says reasoning guidance at inference time raised safe task completion by 33% around obstacles and 43% around nearby people. DeepMind is also releasing ASIMOV-Jailbreak, an open-source benchmark of hundreds of prompts that disguise unsafe requests across dozens of scenes. The 99.5% jailbreak resistance it reports, against 82% for Gemini Robotics ER 2, is for Gemini 4 Argon, a separate model, not for Gemini Robotics 2.5.
Sources
- Google DeepMind, blog post"Gemini Robotics 2.5 accelerates real-world general dexterity," dated Nov 10, 2026
- Google DeepMind, Gemini Robotics pagePartners section: lists Agile Robots, Apptronik and Boston Dynamics as research partners
Found a mistake? See how corrections work.