PHYSICAL AI
Gemini Robotics 2.5 learns tongs, pipettes and Mercedes parts sorting
November 10, 2026 · Updated November 16, 2026
#PhysicalAI #ProductLaunch #Manufacturing #Research
Robots that can tie a bow on a gift box are one step closer to being useful at the holidays, and on the factory floor too. Google DeepMind introduced Gemini Robotics 2.5 on November 10, 2026. It calls the model its most capable vision-language-action system to date. In plain terms, it looks at the scene, listens to instructions and turns both into hand and arm movements.
DeepMind tested the model on more than 200 tasks. The list includes serving food with tongs, moving liquids with a pipette, handling flexible materials and generally being gentle with fragile things. The company reports success rates 30 to 100 percent higher than its earlier baselines, depending on the task, and says jobs finish 20 to 30 percent faster. It also used its Gemini 4 Argon language model as a simulation engine during development, so the robots got plenty of practice without breaking anything real.
The headline demo took place on live Mercedes-Benz production lines in Hungary and Germany. Apptronik's Apollo 2 humanoid, driven by the new model, sorted and kitted automotive parts with a 96 percent success rate. Reports note that the figure covers that specific parts-preparation task, not car-making as a whole, and that no trial count was published. Even so, it is a respectable result for a first shift.
The fine print
- The model: Gemini Robotics 2.5, a vision-language-action model from Google DeepMind
- Announced: November 10, 2026
- Test scope: More than 200 tasks, from tongs to pipettes
- Reported gains: 30 to 100 percent higher success and 20 to 30 percent faster completion versus prior baselines
- Factory pilot: Apollo 2 humanoid at Mercedes-Benz lines in Hungary and Germany, 96 percent success on parts sorting and kitting
- Availability: Rolling out to robotics partners, including Apptronik and Agile Robots, not a public API
What it means for automation buyers
You can't buy this off the shelf yet. Access runs through DeepMind's robotics partners, so the practical route is to ask your robot vendor whether it is in the program. The 96 percent result is promising for kitting and parts preparation, but it comes from the company's own pilot, and a production line cares about the last four percent. If you're planning cobot cells for sorting, packing or delicate handling, treat this as a sign that flexible, instruction-driven manipulation is improving. Then ask vendors for independent cycle-time, uptime and safety data before changing a single purchase order. The company has also released a benchmark for testing robots against malicious prompts, which is worth raising in any future safety review.