Gemini Robotics 2.5 sorts car parts in Mercedes-Benz pilot

Google DeepMind has introduced Gemini Robotics 2.5, a new model for controlling robots, and tested it on active production lines together with the robotics maker Apptronik and Mercedes-Benz. The humanoid robot Apollo 2 sorted metal parts, and Google DeepMind reports a 96 percent success rate for it. The figures come from the announcement of 10 November 2026, which The Decoder covered.

According to The Decoder, Gemini Robotics 2.5 is a so-called vision-language-action model (VLA). Such models process visual information, understand spoken or written instructions and turn them into robot movements. The model builds on a predecessor introduced earlier this year, which was meant to give robots full-body control for the first time. Google DeepMind calls it its most capable model of this kind to date.

What the model is meant to do

According to Google DeepMind, Gemini Robotics 2.5 was tested on more than 200 different tasks, such as serving food with tongs, tying a bow around a gift box or pipetting liquids. Depending on the type of task, the success rate is said to be 30 to 100 percent higher than the company's own previous baselines, and tasks are said to be completed 20 to 30 percent faster.

Caution is in order here: the analysis by the blog CellCog points out that the announcement neither names nor quantifies the comparison baselines. The absolute success rates are therefore unknown, and the gain of 30 to 100 percent cannot be put into context.

The model is meant for different types of robots. The ones named are Apptronik's humanoid robot Apollo 2, whose five-fingered hand has 22 degrees of freedom, and the Franka Duo platform with parallel grippers that have two fingers each.

The pilot at Mercedes-Benz

The practical test took place in Kecskemét in Hungary and in Sindelfingen in Germany, on production lines that are in operation. There the robot handled so-called kitting: it took individual metal parts from bulk bins and placed them in trays in exactly the order in which they are needed for each vehicle on the line. Several components worked together:

• Gemini Robotics ER monitored the workflow.
• Gemini Robotics 2.5 controlled the robot's hands and arms.
• Apptronik's Apollo 2 carried out the grasping and placing movements.

Google DeepMind reports a success rate of 96 percent for the pilot. According to CellCog, the company itself calls it a proof of concept and does not say how many picks the rate is based on. Apptronik chief executive Jeff Cardenas commented in the announcement on the transition from training to use in the plant:

We saw the models improve dramatically in just a few weeks of post-training and carry over to Mercedes-Benz production with very little drop-off.

Safety and simulation

According to Google DeepMind, Gemini Robotics 2.5 uses real-time control that lets the robot avoid physical hazards while performing a task. In its own evaluations, safe task completion improved by 33 percent near obstacles and by 43 percent near people. There is no independent check of these values.

The company is also publishing the open-source benchmark ASIMOV-Jailbreak on Hugging Face. It contains hundreds of prompts that disguise unsafe requests across dozens of scenes, and is meant to test whether robotics models can be tricked into dangerous actions. The language model Gemini 4 Argon reaches a jailbreak resistance of 99.5 percent, while its predecessor Gemini Robotics ER 2 reaches 82 percent. Important: according to CellCog, this score belongs to Gemini 4 Argon, not to Gemini Robotics 2.5 itself; for the robotics model the announcement gives no score.

Argon also played a role in training, as a simulation engine. From a real camera scene it can generate thousands of simulated training runs and build a digital twin. Combined with real demonstrations, this is meant to help the robot generalize faster to new scenes and objects. Argon can also derive hypotheses about the causes of errors from video recordings, which are then tested in simulation and on the real robot.

What remains open

According to Google DeepMind, Gemini Robotics 2.5 is being rolled out to the partners Agile, Apptronik and Mercedes-Benz. According to CellCog, no prices, no public interface and no date for broader access are given, and a technical paper on the model is also missing. All the figures come from Google DeepMind or its partners.

It is also unclear whether the robot stays on the line permanently after the test. A pilot is a trial; lasting use in shift operation would be a different claim. Sensible things to watch are results from independent bodies, measurements with ASIMOV-Jailbreak by other labs, and access for developers outside the partner group.

What this changes for companies in the UK

For suppliers and machine builders among British mid-sized manufacturers, the news is above all a signal, not a product: kitting, meaning the supply of parts in the right sequence, is a task found in many factories, and that is exactly where a robot with a general-purpose control model is now being tried. Gemini Robotics 2.5 cannot currently be bought or deployed. Anyone planning automation should follow the development and ask vendors for concrete details: the sample size behind the success rate, how failed grasps are handled, behaviour near people, and the effort needed to teach the robot new parts.

Until such answers are available, the position is this: the 96 percent is a vendor figure from a pilot, not a proven standard for everyday manufacturing.