Gemini Robotics is the model that turns instructions and camera images into robot movements. According to DeepMind, version 2.5 was tested on more than 200 tasks, such as serving food with tongs and using a pipette, and is also 20%–30% faster. Apptronik's Apollo 2 humanoid learned with it to tie a ribbon into a bow, a particularly hard task for robots. In Mercedes-Benz factories in Europe the robots kitted car parts with a 96% success rate, and a DeepMind director said closing the last 4% will be a "steep road".
Training was sped up with Gemini 4 Argon, which generated thousands of simulated variations of practice runs from real footage, alongside demonstrations by human operators. The model still runs on cloud servers rather than on the robot itself, so there is latency. When something blocks its path it stops or waits, and according to DeepMind this raised safe task success by 33% with obstacles and by 43% when people were in the way.
ASIMOV-Jailbreak is available on Hugging Face. According to DeepMind, with Argon supervising the robot the system scores 98.8% on its safety measure, compared with 73.5% with Gemini Robotics ER 2 alone. Gemini Robotics 2.5 is not publicly available yet; it has gone to partners including Agile Robots, Apptronik and Boston Dynamics.