On November 10, 2026, Google DeepMind released Gemini Robotics 2.5, calling it the most capable vision-language-action (VLA) model it has built: robots can learn more skills, learn them faster, and do them with better dexterity, reliability, and safety. Across more than 200 distinct tasks, it beat previous baselines with success rates 30% to 100% higher and finished tasks 20% to 30% faster. This is not a lab benchmark refresh — the robot went straight onto a car production line.
The clearest evidence comes from a proof of concept with Apptronik and Mercedes-Benz: Apollo 2 humanoids worked on live production lines in Kecskemét, Hungary, and Sindelfingen, Germany, doing automotive kitting on their own — picking metal brackets out of bulk bins and placing them into sequencing containers in the order each vehicle needs, with a 96% success rate. The model splits the job in two: the embodied reasoning model, Gemini Robotics ER, tracks the workflow, while the Gemini Robotics 2.5 action model drives the arms and hands. The research itself got a boost from a new model: Gemini 4 Argon acted as a simulation engine, generating thousands of varied practice runs from a real camera scene so 2.5 could generalize faster to new scenes and objects.
The other side of 96%: you cannot buy it yet
DeepMind is handing the model to its robotics partners only. There is no public API and no announced date for general availability. And 96% is a proof-of-concept score, not a production promise: the remaining 4% of errors on a live line means stoppages and rework. The safety work deserves the same attention: inference-time reasoning guidance lifted safe task completion by 33% around obstacles and 43% around nearby people, and the new open ASIMOV-Jailbreak benchmark tests robots with hundreds of disguised unsafe requests — Gemini 4 Argon resisted 99.5% of them, far above the 82% of the previous baseline, Gemini Robotics ER 2.
The race to put robots in factories is speeding up, and Nvidia's GR00T 2 is chasing the same general humanoid skills. Nvidia's GR00T 2 Is Official: Robots Learn Fine Motor Skills From Human Videos The routes differ, but the goal is the same: get robots steady on a controlled production line first, then take them into the messier real world. For Gemini Robotics 2.5, the next question is whether its partners can turn 96% into everyday routine.
