Google DeepMind’s new robot AI packed car parts with a 96% success rate in a test at Mercedes-Benz factories
Gemini Robotics 2.5 is rolling out to Google’s robot-maker partners, not to the public. The factory work was a trial, not regular production.
Google DeepMind, Google’s AI research lab, on Tuesday introduced Gemini Robotics 2.5, a new AI model that lets robots learn more skills, faster, and work more reliably outside the lab, the lab said in a Nov. 10 announcement.
The model is what DeepMind calls a vision-language-action model. It takes in what a robot’s cameras see and instructions in plain language, and turns them into movements of the robot’s arms and hands. It is the robot equivalent of the AI models behind chatbots, which turn words into words. Robots in factories today mostly repeat movements that engineers program in advance; this kind of model is meant to let a robot handle objects and situations it was not specifically programmed for.
What DeepMind says it can do
DeepMind tested the model on more than 200 tasks, such as serving food with tongs, tying a bow on a gift box and using a pipette, a lab tool for moving small amounts of liquid. Compared with its earlier models, it succeeded 30% to 100% more often, depending on the task, and finished 20% to 30% faster, DeepMind says. It works on different robots, including a humanoid with five-fingered hands and a two-armed robot with simple grippers.
The factory test
In a trial with humanoid robot maker Apptronik and carmaker Mercedes-Benz, Apptronik’s Apollo 2 robots used the model on working assembly lines in Kecskemét, Hungary, and Sindelfingen, Germany. The job, called kitting, means picking metal brackets out of bulk bins and placing them in trays in the exact order each car needs. The robots succeeded 96% of the time, DeepMind says. DeepMind called it a proof of concept, meaning an early test, not a deployment.
Safety
DeepMind says the model can steer around hazards in real time, which raised safe task completion around obstacles by 33% and around nearby people by 43% in its tests. It is also releasing a free test, ASIMOV-Jailbreak, with hundreds of disguised unsafe requests, to check whether robots can be tricked into dangerous actions. Its Gemini 4 Argon model resisted 99.5% of those tricks, DeepMind says.
The model is rolling out to DeepMind’s robotics partners. DeepMind did not say when, or whether, robots using it will be sold to businesses or households.