Google DeepMind has released Gemini Robotics 2.5, a new generation of its physical AI platform designed to perform complex real-world tasks. The update includes Gemini Embodied Reasoning, a vision-language model that uses live video to plan tasks, and Gemini Robotics 2.5, a vision-language-action model that generates physical movements. In laboratory testing, the new model demonstrated a 30 to 100 percent higher success rate and was 20 to 30 percent faster than the previous version.
The platform's capabilities were tested on over 200 activities, including laboratory tasks like using a pipette and serving food with tongs. In one demonstration, Apptronik’s Apollo 2 humanoid robot used its five-fingered hands to tie a bow, a task involving difficult-to-manipulate deformable objects. While the laboratory results were significant, the company also tested the technology in industrial settings through partnerships with Apptronik and Mercedes-Benz.
During these real-world trials, Gemini-powered Apollo 2 robots kitted automotive parts in several European factories, achieving a 96 percent success rate. Despite this performance, the technology has not yet reached the near-perfect consistency of human workers in repetitive manufacturing roles. DeepMind director Kanishka Rao noted that closing the remaining 4 percent gap will be a "steep road."
