AI News

Google DeepMind Releases Gemini Robotics 2.5, Achieving a 96% Sorting Success Rate at Mercedes-Benz Factory

HuxiuUpdated Nov 11 16:30

AI summary

On November 10, Google DeepMind released Gemini Robotics 2.5, calling it its strongest robot vision-language-action (VLA) model to date, with a relative success rate improvement of 30% to 100% across more than 200 tasks and a 20% to 30% increase in execution speed. In proof-of-concept trials on Mercedes-Benz production lines in Germany and Hungary, the Apollo 2 humanoid robot equipped with the model completed parts sorting, grasping, and placement with a task success rate of 96%. The article also mentions that Gemini 4 Argon has been used to generate simulated training scenarios and analyze failure causes, and compares Claude Fable 5.2's 64.9% success rate on 42 tasks in the RoboDojo Sim benchmark with GPT-6 Astra's 29.8%.

Why read this

AI score 75

The article compares Google's new model with the debate over the general-purpose large model approach, and points out the impact of tech giants raising the bar for startups.

Topics

Tags

Body · AI translation

[[**Google DeepMind releases Gemini Robotics 2.5, significantly improving success rates across more than 200 tasks and in a Mercedes-Benz factory proof of concept, proving that specialized embodied models still have value, though the competitive threshold for startups has been raised.**

**Key points**

**1. Factory validation is effective, but large-scale deployment remains far off**

Success rates on more than 200 tasks improved by 30% to 100% relative to the previous baseline, and execution speed increased by 20% to 30%; in the Mercedes-Benz proof of concept, Apollo 2 achieved a 96% success rate in sorting, but continuous operation, human takeover, and other details were not disclosed.

**2. The rise of general models does not negate specialized VLAs**

Claude Fable 5.2 achieved a 64.9% success rate on 42 tasks in RoboDojo Sim, far exceeding GPT-6 Astra's 29.8%, triggering doubts about whether specialized models will be eliminated; however, simulation results are not equivalent to real-world industrial reliability, and physical interaction still requires specialized models.

**3. General models are beginning to train robots**

Gemini 4 Argon can generate simulation scenarios, analyze failures, and help improve training, but still requires real-robot data; competition may shift from data scale to data utilization, automatic problem discovery, and simulation acceleration, where tech giants have greater advantages.

**4. The giants' approaches are converging and raising the barrier for startups**

Both Google and Nvidia adopt general reasoning plus specialized control; GR00T 2 achieved a 95% success rate on 50 unseen tasks after 5 demonstrations and a small amount of fine-tuning. Startups need to prove that their in-house VLAs outperform the giants' foundation models, but Gemini Robotics 2.5 has not fully opened its API, leaving room. Original 2026-11-11 15:09 Embodied models just declared "dead"—saved by Google? Yi Yuqin Quick overview

Just the day before, Anthropic's release of Claude Fable 5.2 prompted many robotics researchers to lament: are specially trained embodied models about to be eliminated by general large models?

In the RoboDojo Sim benchmark, which tests robot manipulation capabilities, Claude Fable 5.2 achieved an average success rate of 64.9% across 42 tasks, far exceeding GPT-6 Astra's 29.8%, and also outperforming most robot-specific models.

More surprisingly, without specialized training for some tasks, Claude was already able to control real robots to twist off bottle caps, solve a Rubik's cube, and even draw.

As general models begin to demonstrate such strong robot manipulation capabilities, the value of specialized embodied models is also being questioned.

However, just one day later, Google gave a different answer.

On November 10, Google DeepMind released Gemini Robotics 2.5, billed as its strongest robot vision-language-action (VLA) model to date.

According to data published by Google, the new model was evaluated on more than 200 tasks; compared with the previous baseline, success rates on some tasks improved by 30% to 100% relative to the baseline, and execution speed increased by 20% to 30%.

Moreover, this model has already entered automotive factories. On Mercedes-Benz production lines in Germany and Hungary, the Apollo 2 humanoid robot equipped with Gemini Robotics 2.5 can autonomously complete the sorting, grasping, and placement of automotive parts, with a task success rate of 96%.

Has the embodied model that was just questioned really been saved by Google?

Google has put embodied large models into factories

This time, the results Google presented are closer to real industrial needs than before.

According to official disclosures, Gemini Robotics 2.5 mainly achieves improvements in three areas: dexterous manipulation, task reliability, and physical safety.

First is manipulation capability.

Google tested the new model on more than 200 tasks, covering operations such as using tongs to pick up food, tying a bow on a gift box, and using a pipette.

These tasks may look simple, but they are not easy for robots.

Take tying a bow as an example: the robot needs to control both hands simultaneously, perceive the shape of the ribbon, and continuously adjust force and motion according to how the material deforms. With the slightest deviation, the ribbon may slip or the knot may fail.

In tests published by Google, the new model increased the success rate of using tongs to pick up pastries from 5% to 100%, increased the success rate of untying a bow from 0 to 70%, and increased the success rate of closing a toolbox lid from 0 to 93%.

Overall, compared with the previous baseline model, Gemini Robotics 2.5 improved success rates on different categories of tasks by 30% to 100% relative to the baseline, and increased execution speed by 20% to 30%.

At the same time, the new model can adapt to different types of robots, including Apollo 2 equipped with five-fingered dexterous hands and Franka Duo using two-finger parallel grippers.

This means Google is trying to make the same set of robot intelligence capabilities applicable to different hardware forms, reducing the cost of retraining the model every time the robot body changes.

But the more noteworthy progress took place in the Mercedes-Benz factory.

On automotive production lines in Sindelfingen, Germany, and Kecskemét, Hungary, Google collaborated with Apptronik and Mercedes-Benz to have the Apollo 2 humanoid robot handle the sorting and kitting of automotive parts.

Specifically, the robot needs to take parts from bins containing metal brackets and then place them into designated trays in production order, preparing for subsequent automobile assembly.

This kind of work is highly repetitive, yet it requires the robot to accurately identify, grasp, and place parts. Once a deviation occurs, it also needs to be corrected in time.

Google adopted an approach in which two sets of models work in coordination: one set of models understands the task and the workflow, while the other is responsible for controlling the robot's arms and dexterous hands to complete the specific operations.

Ultimately, Apollo 2 achieved a 96% task success rate in this proof of concept.

Jeff Cardenas, co-founder and CEO of Apptronik, said that with Gemini Robotics, Apollo 2 made significant progress in just a few weeks of post-training, and after being transferred to Mercedes-Benz's real production line, its performance did not show any notable decline.

However, a 96% success rate is still a gap away from large-scale industrial application.

Google has not yet disclosed the complete number of tests, the duration of continuous operation, the frequency of human intervention, or the difference in production takt time between the robot and human workers.

For an automobile factory, completing a single sorting task and working continuously and stably for thousands of hours are two different things.

Especially on industrial production lines, what truly determines whether a robot can stay is not only the operation success rate, but also fault recovery capability, maintenance cost, and return on investment.

Therefore, Apollo 2's performance is more like a rather convincing technical validation, and there are still many questions to be answered before large-scale commercial deployment.

But at least Google has proven that the capabilities of embodied models can extend from the laboratory to real industrial environments.

Compared with how many new actions the robot has learned, there is another change in this release that is easy to overlook: Google is changing the way embodied models are trained.

In the past, when robots learned new skills, researchers often needed to repeatedly collect data, design training tasks, and then continuously revise them through real-machine testing.

This process was both expensive and time-consuming.

Especially when dexterous manipulation is involved, the robot needs to learn how to control grasp force, handle object deformation, and cope with the complex changes produced during contact.

Even if the model makes progress on a certain task, changing to another object or another scenario may still require retraining.

Now, Google is trying to let general foundation models take on part of the work.

According to disclosures from Google DeepMind, Gemini 4 Argon has been introduced into the robot R&D process to generate simulated training scenarios, analyze the causes of failures, and assist researchers in improving training strategies.

For example, researchers only need to provide camera images of a real scenario, and Argon can help build a digital twin environment and generate thousands of simulated practice runs under different conditions.

If the robot fails during a certain operation, the model can also analyze the failure video, propose possible causes, and assist researchers in verifying improvement plans in simulation or real-machine environments.

English text is being prepared.

But its potential value is very obvious.

In the past, for a robot to master a new skill, a large amount of human-involved trial and error was required. With the intervention of general models, part of the training work can be completed in virtual environments, and researchers can also find effective directions for improvement more quickly.

This means that competition in embodied models may undergo new changes.

Previously, the industry generally focused on who had more robot data and who had established larger-scale data collection centers.

In the future, in addition to data scale, whether a model can use data efficiently, automatically discover problems, and accelerate learning through simulation will also become important competitive indicators.

And in this regard, tech giants with strong general foundation models, computing power, and simulation technology reserves clearly have certain advantages.

For Google, Gemini 4 Argon and Gemini Robotics 2.5 are forming a mutually coordinated R&D system.

General models can assist robot learning, while specialized models continuously accumulate real-world manipulation capabilities.

This combination also provides a new perspective for observing the recent debate over embodied model routes.

Has the embodied model been saved by Google?

Over the past two years, there have been two different technical routes in the embodied intelligence industry.

One is to specially train robot models, enabling robots to master manipulation capabilities through large amounts of real-machine data, motion data, and physical interaction experience.

The other is to use the visual understanding, reasoning, and planning capabilities already possessed by general foundation models and further extend them to robot control.

As general models such as GPT-6 Astra and Claude Fable 5.2 continue to make breakthroughs in robot tasks, the second route has begun to attract more attention.

The logic behind it is not complicated.

If general models can already understand the environment, plan actions, and even control robots to complete some tasks, then is it still necessary for companies to invest large amounts of money to train an embodied foundation model from scratch?

However, Google's latest progress shows that the issue may not be that simple.

Although general models have demonstrated relatively strong robot manipulation capabilities, there is still a gap between understanding a task and stably completing an operation in the real world.

For example, the model can accurately determine how a cup should be grasped, but during actual execution it still needs to handle issues such as contact between the fingers and the cup, grasp force, robotic arm movement, and object slipping.

This kind of continuous physical interaction places higher demands on control frequency, motion precision, and environmental feedback.

General foundation models are still unable to independently solve all problems at present.

It is worth noting that even if Anthropic's Claude Fable 5.2 achieved a 64.9% success rate in the RoboDojo Sim test, that result mainly reflects manipulation capability under specific simulation conditions and cannot be directly equated with reliability in real industrial environments.

Therefore, Google still chose to retain a dedicated VLA model, controlling the robot through collaboration with a reasoning model.

NVIDIA's newly released GR00T 2 also adopts a similar approach.

In the system it demonstrated, the general reasoning model is responsible for decomposing tasks, planning steps, and checking progress, while GR00T 2 handles the specific physical motion control.

Among 50 tasks that the target robot had never seen before, after 5 demonstrations per task and a small amount of fine-tuning, GR00T 2 achieved an average success rate of 95%.

The two giants are exploring a similar direction: letting general models take on more understanding and reasoning work, while dedicated robot models handle physical operations.

This means that embodied models still have clear technical value in the short term.

But for startups dedicated to developing embodied models, this is not necessarily good news.

When giants such as Google and NVIDIA simultaneously possess foundation models, robot models, and related training tools, the competitive threshold for startups may be further raised.

In particular, companies that mainly rely on self-developed VLAs to build technical barriers will need to answer a sharper question in the future: compared with directly using the foundation models provided by the giants, what exactly is better about the models they train themselves?

Of course, the giants have not yet fully mastered the robot market.

At present, Gemini Robotics 2.5 is mainly opened to Google's robot partners and has not yet fully provided public API services. The scenario requirements, hardware configurations, and data conditions of industrial customers also vary greatly, and general models are still a long way from full adaptation.

This leaves room for development for specialized embodied model companies.

Looking back, Google's release this time may have provided a provisional answer to the recent debate.

Embodied models are still necessary, but in the future not every company will necessarily need to train a set from scratch.

As general foundation models continue to expand the boundaries of their capabilities, robot-specific models will also face stricter value tests.

Google may have saved embodied models, but it has left some embodied model startups facing more brutal competition.

Article title: The embodied model that was just sentenced to "death penalty"—saved by Google?

Has published 16 articles on Huxiu. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]

Source

Huxiu