Gemini Robotics 2.5: Humanoid robot sorts Mercedes parts with a 96% success rate

What's going on

On November 9, 2026, Google DeepMind unveiled Gemini Robotics 2.5. The AI model controls robots, and its main job is to give them a finer touch. DeepMind calls it its most capable vision-language-action model (VLA) to date. That's an AI that sees camera images, understands instructions in everyday language and turns them directly into movement commands for arms, hands and grippers.

The announcement focuses on three things: dexterity, reliability and safety. As proof of reliability, DeepMind points to a pilot project with robot maker Apptronik and Mercedes-Benz, where Apollo 2 humanoid robots worked on live production lines. For now, the model only goes to Google's robotics partners. There's no public API and no download.

Gemini Robotics 2.5 builds on the Gemini Robotics 2 family, which Crypto Briefing reports was introduced on July 30, 2026. That release was about whole-body control for humanoid robots. Version 2.5 adds finer manual work and faster learning of new skills.

Two AI layers: planning and acting

As Ars Technica explains, DeepMind's robotics platform is made up of several layers. At the top sits Gemini Robotics ER (ER stands for embodied reasoning, meaning reasoning tied to a body in the real world). This model takes the instruction, watches the scene via live video and plans the steps. Below it runs the actual VLA model, Gemini Robotics, now at version 2.5. It generates the concrete movements, such as how a joint or an actuator (a motor or drive element in the robot) should move.

Put simply: ER is the foreman keeping track of the workflow, and Gemini Robotics 2.5 is the hands. According to DeepMind, the model works with different types of robots. These include Apptronik's Apollo 2 humanoid with the five-fingered SharpaWave hand and its 22 degrees of freedom (22 independently movable joint axes), as well as the Franka Duo platform with two simple parallel grippers.

The ER layer, by the way, is already available to developers. According to Technobezz, Gemini Robotics ER 2 has been available through Google AI Studio and the Gemini API since July, and is now also in public preview in Google Cloud's Model Garden. There, you can hook up motion systems such as VLA models as tools. ER 2 is meant to let a robot check whether a step worked and repeat only that step. For spotting the decisive video frame, say, when to stop pouring, Google cites an accuracy of 91.3 percent.

What the robot can do now

DeepMind tested Gemini Robotics 2.5 on more than 200 different tasks. Examples include serving food with tongs, tying a bow around a gift and pipetting liquid. These tasks demand tool use, handling fragile items and, above all, dealing with deformable objects like fabric or ribbon. Ars Technica notes that such objects are notoriously hard for robots because they constantly change shape. The magazine calls it a remarkable feat that Apollo 2 can tie a bow with its five-fingered hands. Crypto Briefing also lists folding laundry among the skills shown.

The numbers, according to the company:

Ars Technica points out that the gains in success rate and speed come from lab tests. There are no independent measurements yet.

The real-world test at Mercedes-Benz

The most tangible part of the announcement is the pilot at the Mercedes plants in Kecskemét (Hungary) and Sindelfingen. There, Apollo 2 took over what's known as kitting: individual parts are picked from bulk bins and sorted into containers in exactly the order they'll later be needed on the line for a specific vehicle. The work is monotonous and physically demanding.

Gemini Robotics ER kept track of the workflow, while Gemini Robotics 2.5 controlled the arms and hands. The robot picked metal brackets from the bins, placed them in the sequencing containers and filled gaps on its own. It succeeded 96 percent of the time. DeepMind says this should free up technicians for more complex assembly work.

Robotics commentator Lukas M. Ziegler, writing on LinkedIn (quoted by humanoid.guide), relays an assessment from Apptronik CEO Jeff Cardenas: a few weeks of post-training significantly improved the models, and performance barely dropped when moving into real Mercedes production. humanoid.guide also puts the trial in perspective: it's a clearly defined pick-and-sort workflow, not an autonomously running factory.

Ars Technica takes a critical view of the number. Compared with earlier robot AIs, 96 percent is impressive, but for a company like Mercedes-Benz it's probably not good enough yet. A human gets a routine task like this right virtually every time. DeepMind director Kanishka Rao himself says closing the last four percent is

“a steep climb.”

Safety: obstacles, people and jailbreaks

Once robots work in homes, shops or factory halls, physical safety becomes critical. Gemini Robotics 2.5 handles this with runtime reasoning control. While the robot works, an extra reasoning layer steers it around hazards, similar to the safety filters on Gemini chat responses. In DeepMind's tests, this boosted safe task completion by 33 percent around obstacles and by 43 percent near people.

Also new is ASIMOV-Jailbreak, an open-source benchmark. A jailbreak is an attempt to trick an AI into bypassing its safety rules with cleverly worded instructions. The benchmark contains hundreds of prompts that disguise dangerous requests across dozens of scenes. According to DeepMind, its large model Gemini 4 Argon resists in 99.5 percent of cases. The previous baseline, Gemini Robotics ER 2, scored 82 percent. DeepMind says Argon also leads on other agent tests: refusing unsafe tasks, handling ambiguity and working close to people.

Gemini 4 Argon as a training accelerator

Teaching robots new skills usually takes huge amounts of real-world motion data, and that's expensive. So DeepMind uses Gemini 4 Argon as a kind of simulation engine. From a single real camera recording, Argon can build a digital twin, a realistic virtual copy of the scene, and generate thousands of varied practice runs inside it. Combined with real demonstrations, this helps the robot learn to handle new scenes and objects faster. Argon also helps researchers figure out from videos why an attempt failed, and test possible fixes in simulation and on the real robot.

Our take

Gemini Robotics 2.5 isn't a product you can try out. It's a research and partner update. Still, it's worth paying attention to. The bow and the pipette show that fine motor skills with deformable objects are making progress. And with the Mercedes pilot, the technology leaves the lab and earns a concrete success rate from a real factory.

Caution is still warranted. Nearly all the figures come from DeepMind itself. The 30 to 100 percent improvements are measured against unspecified earlier models and come from lab tests. The factory deployment covers a single, narrowly defined task. And as Ars Technica stresses, 96 percent usually isn't enough in industry when a human works almost flawlessly. On the plus side, DeepMind is releasing ASIMOV-Jailbreak as an open tool that other manufacturers can use to test their robot AIs against manipulation. In robotics, after all, the stakes aren't just inappropriate text, but real physical harm.