Google DeepMind has unveiled Gemini Robotics 2.5, an advanced vision language action (VLA) model designed to control robots with greater precision and efficiency. The model recently achieved a 96 percent success rate in a pilot test at Mercedes-Benz production lines in Hungary and Germany, handling kitting tasks with humanoid robots.
Gemini Robotics 2.5 builds on the capabilities of its predecessor, introduced earlier this year, which was the first to provide full body control to robots. The new model has been tested on over 200 tasks, including serving food with tongs and pipetting liquids, outperforming the company’s previous baselines by up to 100 percent in some cases and completing tasks 20 to 30 percent faster.
The model is versatile, capable of working with various robot types. It was tested alongside Apptronik’s Apollo 2, a humanoid robot with a five fingered hand that offers 22 degrees of freedom, and the Franka Duo platform, which features two finger parallel grippers. Together with Apptronik and Mercedes-Benz, DeepMind conducted the pilot test at active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The Apollo 2 robot was tasked with sorting metal parts from bulk bins and placing them in trays, aligning with the assembly line needs of each vehicle.
During the pilot, two AI models collaborated: Gemini Robotics ER monitored the workflow, while Gemini Robotics 2.5 controlled the robot’s hands and arms. According to Jeff Cardenas, CEO of Apptronik, the models improved significantly within a few weeks of post training and successfully transitioned to Mercedes production with minimal performance loss.
Gemini Robotics 2.5 is designed to operate in real time, ensuring that robots can avoid physical hazards while performing tasks. Evaluations show that safe task completion improved by 33 percent near obstacles and by 43 percent near people. To test the model’s robustness, Google DeepMind published the open source ASIMOV Jailbreak benchmark on Hugging Face. This benchmark contains hundreds of prompts that disguise unsafe requests in various scenes and aims to determine whether the model can be tricked into performing dangerous actions. According to the company, Gemini 4 Argon, DeepMind’s language model used in the development of Gemini Robotics 2.5, achieved a jailbreak resistance of 99.5 percent, compared to 82 percent for its predecessor.
In developing Gemini Robotics 2.5, DeepMind utilized its own Gemini 4 Argon language model as a simulation engine. This model can generate thousands of simulated training runs from real camera scenes and create digital twins, helping the robot generalize more quickly to new scenes and objects. Argon can also derive hypotheses from video recordings and test improvements in both simulation and on the actual robot.
Google DeepMind is currently rolling out Gemini Robotics 2.5 to partners including Agile, Apptronik, and Mercedes-Benz. The company offers subscribers to its THE DECODER newsletter ad free reading, a weekly AI newsletter, exclusive “AI Radar” reports, full archive access, and access to its comment section.
Thinking about building an AI product?
Get in Touch