Google DeepMind has introduced Gemini Robotics 2.5, a robot-control model that combines vision, language and actions. The company says it is designed to help robots handle objects more precisely, including delicate and flexible items. The model is beginning to roll out through DeepMind’s robotics partners.
Learning skills and handling objects
The new version draws on the capabilities of the Gemini 4 Argon language model. DeepMind says these allow robots to learn a broader range of skills more quickly and apply them in real-world situations.
According to DeepMind, improved dexterity is intended to let robots perform more tasks with greater precision. The activities it describes include using everyday tools to handle delicate and flexible objects.
Model evaluations and Mercedes-Benz trial
DeepMind tested the model on more than 200 distinct tasks and expects to scale it further. In a real-world proof-of-concept test at Mercedes-Benz, it achieved a 96% success rate packing automotive parts, according to DeepMind.
In other evaluations, DeepMind says success rates ranged from 30% to 100%, depending on the task. The model completed those tasks 20% to 30% faster than the previous version used as a baseline.
Apptronik co-founder and CEO Jeff Cardenas said models paired with the Apollo 2 humanoid improved substantially over a few weeks of post-training. He said their performance dropped only slightly when transferred to Mercedes-Benz production.
Supported robots and availability
Like its predecessors, the model works natively with different types of robots, according to DeepMind. These include Apptronik’s Apollo 2 humanoid and the Franka Duo platform, which has two-fingered parallel grippers.
DeepMind also reports better control over robots’ physical movements and has introduced a benchmark for testing jailbreak attempts to bypass the model’s restrictions. Partners through which Gemini Robotics 2.5 is beginning to roll out include Agile Robots, Apptronik and Boston Dynamics.
