What changed
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model that achieved a 96% success rate in sorting metal parts on active Mercedes-Benz production lines. The model, which builds on its predecessor, was tested in tandem with Gemini Robotics ER, with the latter monitoring workflow and the former controlling hands and arms. The pilot involved Apptronik's humanoid robot Apollo 2, which sorted parts from bulk bins and placed them into trays in the order needed for each vehicle on the assembly line.
What this means in practice
The model is designed to work with different robot types, including Apptronik's Apollo 2 with a five-fingered hand and the Franka Duo platform. It was tested on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. According to Apptronik CEO Jeff Cardenas, the models improved significantly within weeks of post-training and transitioned to Mercedes production with minimal performance loss. The success rate of 96% highlights the model's potential for industrial automation tasks.
Limitations and availability
Gemini Robotics 2.5 is being rolled out to partners Agile, Apptronik, and Mercedes-Benz. Google DeepMind also published the open-source ASIMOV-Jailbreak benchmark to test the safety of robotics models. The company claims Gemini 4 Argon, used as a simulation engine during development, achieves a 99.5% jailbreak resistance, compared to 82% for its predecessor. The model is intended for use in industrial settings, with no details provided on licensing or API availability.
