96% at Mercedes-Benz: the gap between robot demo and real deployment just shrank
Google's new robot software scored 96% on a live car-factory floor — which means robots are no longer just for demos, they are doing real industrial work.
- 96%
- success rate on live Mercedes-Benz production line
- 200
- distinct tasks tested in evaluation
- +43%
- safer around people than previous version
- 20–30%
- faster task completion
On November 10, Google DeepMind released Gemini Robotics 2.5, a new version of its robot-brain software. Within days, humanoid robots from partner Apptronik were running on live production lines at Mercedes-Benz in Hungary and Germany, sorting car parts at a 96% success rate. That is not a lab test. That is a robot doing real work on an active production line.
The jump from the previous version is big. DeepMind tested the model across 200 different tasks. On several of them, the old version scored zero. The new model completes tasks 20 to 30 percent faster and succeeds much more often — on some dexterous tasks, the success rate went from near-zero to over 90 percent. That kind of leap in one release is rare.
Safety improved significantly too. The model works 43 percent better around people and 33 percent better near obstacles. DeepMind also published ASIMOV-Jailbreak, a public test with hundreds of prompts designed to trick a robot into doing something dangerous. Their best model blocked 99.5 percent of those attempts, up from 82 percent in the previous version.
This is not just one model update. It is a signal. Robots are crossing a line — from compelling demos into real industrial deployments. Manufacturing, logistics, and warehousing will see this first. If your business involves repetitive physical tasks, the gap between demo and real deployment just shrank — and that changes what you should be planning now.