SIAgentNews.com
← All news
DevelopingGeminiSecurity

Google DeepMind tests Gemini Robotics 2.5 in Mercedes-Benz factories

The new vision-language-action model achieved a 96% success rate in automotive parts-handling tasks using a humanoid robot.

· First reported by AI Insider

96%Success rate in factory tests

The short version

  • Gemini Robotics 2.5 achieved a 96% success rate in automotive kitting tasks at Mercedes-Benz factories.
  • The model completed tasks 20% to 30% faster than previous baselines across more than 200 different evaluations.
  • Success rates for specific complex tasks, like closing a toolbox lid, rose from 0% to 93%.
  • Google DeepMind introduced ASIMOV-Jailbreak to test if robots can resist disguised unsafe instructions.
  • The model is being rolled out to partners including Boston Dynamics and Agile Robots.

Why it matters

This move signals a shift from robots performing repetitive motions in cages to humanoids that can handle diverse, delicate objects in shared workspaces. It suggests that AI can now bridge the gap between digital reasoning and complex physical dexterity.

Factory floor testing

Google DeepMind recently debuted Gemini Robotics 2.5, a vision-language-action model designed to help robots learn physical tasks more quickly. The company tested the model using Apptronik’s Apollo 2 humanoid robot on active production lines at Mercedes-Benz factories in Hungary and Germany. These robots performed automotive kitting, which involves sorting individual parts from bulk containers into specific trays for vehicle assembly.

During these tests, the robot had to pick up metal automotive brackets from supply bins and place them into sequencing containers. The system had to be precise enough to correct for gaps as it worked through the assembly process. As AI Insider reported, the model achieved a 96% success rate during these specific automotive parts-handling evaluations.

The demonstration utilized a combination of two different models to manage the workload. Gemini Robotics ER handled the overall workflow management, while Gemini Robotics 2.5 specifically controlled the movements of the Apollo 2's hands and arms. This division of labor allowed the robot to navigate the complexities of a real industrial environment alongside human workers.

Significant dexterity gains

The new model showed substantial improvements over previous versions when handling objects that require high levels of precision. Google DeepMind evaluated the system across more than 200 tasks, including manipulating flexible materials, using tools, and handling delicate items. In many of these categories, the success rates were 30% to 100% higher than previous baselines.

The jump in performance was particularly noticeable in tasks that the previous models struggled with entirely. For example, the success rate for packing pastries with tongs rose from 5% to 100%. Other notable improvements included closing a toolbox lid, which went from 0% to 93%, and untying ribbon bows, which increased from 0% to 70%. The robot also showed progress in hanging a cardigan and sweeping trash into a dustpan.

These results indicate that the model is becoming much better at understanding the physics of different materials. It can now interact with items like liquids, food, and clothing with greater reliability. The system is designed to work across various platforms, including the Apollo 2's five-fingered hand and the Franka Duo system's two-fingered grippers.

Safety and reasoning

Physical safety is a critical hurdle for deploying robots in human-centric environments. Gemini Robotics 2.5 uses reasoning to adjust its movements around obstacles and people in real-time. In company evaluations, this approach led to a 33% increase in safe task completion around obstacles and a 43% increase in safety when operating near people.

To further ensure these systems remain safe, Google DeepMind introduced a new open-source benchmark called ASIMOV-Jailbreak. This benchmark contains hundreds of prompts designed to see if a robotics AI can resist disguised or hidden unsafe instructions. The company reported that the Gemini 4 Argon model achieved a 99.5% resistance rate on this benchmark, a significant improvement over earlier models.

The model was also tested on its ability to resolve uncertainty and refuse tasks that were deemed unsafe. By combining these safety protocols with advanced reasoning, the developers aim to make humanoid robots more viable for commercial use. These features are intended to ensure that robots can operate autonomously without posing risks to the people working alongside them.

Accelerating robot training

A key part of the development process for Gemini Robotics 2.5 involved using another model called Gemini 4 Argon. This model helps researchers generate thousands of simulated practice runs based on real-world camera scenes. By creating these varied conditions, researchers can train and evaluate robots in a virtual space before they ever touch a physical object.

These simulations can be combined with actual demonstrations collected from physical robots to help the AI adapt to unfamiliar environments. If a robot fails a task in the real world, Argon can examine the video of the failure to suggest why it happened. This feedback loop allows researchers to test changes in the simulation and then apply those fixes to the physical robot.

Google DeepMind noted that these capabilities significantly sped up the development of the Gemini Robotics 2.5 model. By using AI to analyze failures and generate training data, the company can move from a concept to a functional factory robot much faster. The model is now being rolled out to several key partners, including Boston Dynamics and Agile Robots, to expand its use in industrial settings.

What to watch

  • How quickly these robots move from factory testing to wider commercial adoption.
  • Whether the 96% success rate holds steady as tasks become more complex or unpredictable.

Try it

Look up the 'ASIMOV-Jailbreak' benchmark to see how researchers are testing the safety of robotics AI.

Sources and links

First reported by AI Insider. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news