Google tests Gemini Robotics 2.5 on automotive parts handling
The AMW Read
The release updates a major player's robotics capabilities and reports a factory test, but the available results do not establish sustained commercial performance.
Named counterparties: Apptronik
Google tests Gemini Robotics 2.5 on automotive parts handling
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model for physical tasks. The company reported a 96% success rate in an automotive kitting test using Apptronik's Apollo 2 humanoid on Mercedes-Benz production lines in Hungary and Germany. Gemini Robotics ER managed the workflow, while the new model controlled the robot's hands and arms. Google says it evaluated the model across more than 200 tasks and is rolling it out to Agile Robots, Apptronik, and Boston Dynamics.
The release advances Google's position in the market for models that can operate across different robot platforms. The reported factory result offers a test of that approach in an industrial setting, while comparisons across tasks suggest improved dexterity. Google says success rates were 30% to 100% higher than earlier baselines, depending on the task. The source does not specify the number of factory attempts or the baseline for the 96% result, which limits comparison with other systems.
For builders and investors, the next question is whether performance holds across parts, shifts, and production sites at a cost customers can justify. Google also reported safety improvements and introduced ASIMOV-Jailbreak, a benchmark for resistance to unsafe instructions. Its reported 99.5% result applies to Gemini 4 Argon on that benchmark, not to the factory test. Neither figure establishes long-term reliability or commercial returns.
