Google DeepMind tests Gemini Robotics 2.5 on Mercedes-Benz production lines
The new AI model achieved a 96% success rate in a parts-sorting proof of concept, signaling a step toward automating complex factory tasks.
The short version
- Gemini Robotics 2.5 achieved a 96% success rate on specific parts-sorting and kitting tasks at Mercedes-Benz facilities.
- The model improved dexterous task performance by 30% to 100% across roughly 200 different evaluations.
- The system uses 'Gemini Robotics ER' to track overall task progress while 2.5 controls specific arm and hand movements.
- The demonstration utilized the Apollo 2 humanoid robot to handle metal brackets in a live manufacturing environment.
- Safety features include hazard avoidance, resistance to dangerous prompts, and hardware-level stopping mechanisms.
Why it matters
This move signals a shift from robots performing repetitive, isolated motions to handling complex, variable tasks in live human environments. For manufacturers, it suggests a future where robots handle tedious parts preparation, allowing human workers to focus on high-level assembly.
Live factory testing in Europe
Google DeepMind recently announced Gemini Robotics 2.5, a specialized AI model designed to govern how robots move and interact with their surroundings. The company conducted a proof of concept using this model on live production lines for Mercedes-Benz in Hungary and Germany. This real-world application marks a significant step in moving robotics from controlled labs to active industrial settings.
During these tests, the system was tasked with a specific parts-preparation job. A humanoid robot had to pick up metal brackets that were tossed loosely into supply bins and arrange them into a different bin in a specific order. The goal is to automate these repetitive tasks so that human workers can dedicate their time to more complex assembly steps that require human judgment.
The results of this specific demonstration were notable, with the system reaching a 96% success rate on the sorting and kitting tasks. However, DeepMind clarified that this figure applies only to this specific parts-preparation task. It does not represent the success rate for the entire automotive manufacturing process or long-term stable operation across a whole factory.
How the AI controls movement
The system operates by combining two distinct AI components to manage the robot's behavior. Gemini Robotics ER is responsible for tracking the progress of the overall task at hand. Meanwhile, Gemini Robotics 2.5 focuses on the granular control of the robot's hands and arms to execute specific physical movements.
To improve the model's capabilities, DeepMind used a method that generates thousands of simulated practice scenarios. These are created from footage of real environments captured by cameras to build virtual copies of actual settings. By merging these simulations with actual demonstration data from real robots, the company aims to help the AI handle a wider variety of unpredictable situations.
The model was also evaluated on roughly 200 different tasks to test its dexterity. These included complex actions like handling food with tongs or tying a ribbon on a gift box. These evaluations showed that success rates rose by 30% to 100% compared to earlier versions, depending on the specific type of task being performed.
Safety and hazard avoidance
Safety in industrial robotics involves three different layers that must be evaluated separately. First, the AI model handles hazard avoidance, which is the ability to complete tasks while avoiding people or obstacles. DeepMind reported that the ability to complete tasks safely improved by 33% around obstacles and 43% around people during their evaluations.
Second, the system includes a layer of resistance to dangerous instructions. This was tested using a public benchmark called ASIMOV-Jailbreak, which measures how well the AI refuses hundreds of prompts designed to disguise dangerous requests. While the underlying Gemini 4 Argon model scored 99.5% on this test, it is a separate metric from the physical safety of the robot's movements.
Finally, the hardware itself includes a physical stopping mechanism. The Apollo 2 robot used in the test features a zone that triggers an immediate stop when an object is detected. These three layers—AI inference, instruction refusal, and hardware stops—work together to ensure the robot operates safely in a shared workspace with humans and equipment.
The path to full deployment
While the Mercedes-Benz demonstration is a major milestone, several hurdles remain before full factory deployment is possible. The current results do not provide data on how often humans had to intervene during the work or how the robot handles recovery from failures. For a robot to be useful in a factory, it must be able to self-correct when a task goes wrong without constant human supervision.
The company is currently providing early access to Gemini Robotics 2.5 for specific partners in the robotics field. Interested parties must verify eligibility and terms of use, as the model is not currently available to the general public. Additionally, the robot's own management systems, such as perception and task coordination, must be integrated with Google's model.
Future success will depend on how performance and work efficiency change during continuous, long-term operation. As more operational data becomes available, it will be easier to determine if these robots can replace human labor in parts preparation. For now, the 96% success rate serves as a proof of concept for high-probability task completion in a live environment.
What to watch
- Data on how often human intervention is required during continuous factory shifts.
- How the model handles 'recovery' actions when it drops or misplaces a part.
- The expansion of the model to support whole-body coordination like walking and squatting.
Try it
Look up the 'ASIMOV-Jailbreak' benchmark to see how researchers test AI's ability to refuse dangerous instructions.
Sources and links
- ReportingXenoSpectrum: the article we sourced
First reported by XenoSpectrum. 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.