During trials at Mercedes-Benz production sites in Germany and Hungary, the Apollo 2 humanoid demonstrated 96% accuracy in sorting automotive parts. Google DeepMind's new development allows machines to not just follow a program, but to independently analyze the environment and make decisions based on visual data and text instructions.
A technological breakthrough in robot control
The Gemini Robotics 2.5 model belongs to the VLA (Vision Language Action) class. This means the system converts camera information and text commands directly into physical actions. Instead of rigidly programmed movements, the robot receives a task, assesses its surroundings, and determines for itself exactly how to move its arms or grasp an object.
Compared to the previous generation of systems, new efficiency metrics have increased significantly. The trials covered more than 200 different scenarios, including working with fragile objects, flexible materials, and tools.
Gemini Robotics 2.5 test results
Context
The trials were conducted in collaboration with Apptronik and automaker Mercedes-Benz. The tests verified the ability of humanoids to perform useful work directly within the conditions of an active factory.
