Google DeepMind releases robot AI, citing 96% success in Mercedes factory pilot

Google DeepMind is putting its latest robot-control model into partners’ hands after a factory pilot that it says successfully sorted and packed automotive parts 96% of the time. Announced November 10, Gemini Robotics 2.5 combines improved manipulation skills with faster task execution, but the Mercedes-Benz trial remains a proof of concept—not a claim of finished industrial reliability.

The release is a vision-language-action model: the part of DeepMind’s robotics system that turns visual information and instructions into physical movements. A separate embodied-reasoning model handles planning. Gemini Robotics 2.5 is rolling out to robotics partners rather than becoming generally available, so this is not an open invitation to download and run it.

Sorting parts on an active production line

DeepMind tested Apptronik’s Apollo 2 humanoid robots at Mercedes-Benz production facilities in Kecskemét, Hungary, and Sindelfingen, Germany. Their job was automotive kitting: taking individual parts from bulk bins and arranging them in trays in the order needed for each vehicle. It is repetitive preparation work, rather than assembling an entire car or taking over a production line.

Two models divided the work. Gemini Robotics ER tracked the workflow, while Robotics 2.5 controlled Apollo’s hands and arms. The robot reached into bins, grasped metal brackets and placed them in sequencing containers. DeepMind’s stated aim is to shift that preparation away from technicians so they can focus on more complex build tasks.

That result still leaves a reliability gap. DeepMind director Kanishka Rao told Ars Technica that closing the remaining four percent would be a “steep road.” Robotics head Carolina Parada described a tension between making a robot excellent at one application and preserving its ability to handle many different jobs.

We can be very good at a particular application at the cost of generalization

Carolina Parada, head of DeepMind Robotics, speaking to Ars Technica

Practice worlds feed the action model

Beyond the factory pilot, DeepMind evaluated the model across roughly 200 tasks, including serving food with tongs, tying a bow and pipetting liquid. These test different kinds of dexterity: handling tools, manipulating flexible materials and moving delicate objects. The company reports success rates 30% to 100% higher than previous baselines, depending on the task, with completion 20% to 30% faster.

Gemini 4 Argon helped produce the training material. DeepMind says it generated thousands of simulated practice runs from real camera scenes, building digital twins—virtual versions of physical settings—with varied objects and conditions. The team combined those simulations with real demonstrations to help robots adapt to unfamiliar scenes. Argon also reviewed failure videos and helped researchers propose and test fixes.

The action model supports different robot designs, including Apollo 2’s five-fingered hands and the Franka Duo’s two-finger grippers. But broader hardware support does not mean computation has moved onto the robot: Ars reports that Robotics 2.5 still runs on cloud servers, leaving latency alongside the reported speed improvements.

Safety covers movement and instructions

The release addresses two distinct safety problems: moving around physical hazards and refusing dangerous requests. For movement, Robotics 2.5 uses reasoning during operation to guide actions around hazards. DeepMind reports improved safe task completion in its evaluations, rather than claiming the robots cannot cause an accident.

• Physical hazards: DeepMind says its guidance increased safe completion by 33% around obstacles and 43% around nearby people.
• Unsafe instructions: The accompanying open-source ASIMOV-Jailbreak benchmark contains prompts that disguise unsafe requests across dozens of scenes, testing whether a robotics system can be talked into dangerous behavior.

On that instruction-level test, DeepMind reports 99.5% jailbreak resistance for Argon, versus 82% for its previous Gemini Robotics ER 2 baseline. Those are results for Argon’s resistance to adversarial requests—not the factory pilot’s task-success rate or a measurement of collision-free movement.