DeepMind unveils Gemini Robotics 2.5 for more capable robots
DeepMind has introduced Gemini Robotics 2.5, a vision-language-action model designed to help robots learn a wider range of physical skills more quickly. The company evaluated it on more than 200 tasks, including serving food with tongs, tying a gift-box bow and pipetting liquid. Compared with previous baselines, DeepMind reports success-rate improvements of 30% to 100%, depending on the task, while completion times fell by 20% to 30%.
The model works with multiple robot types, including Apptronik’s Apollo 2 humanoid and the Franka Duo platform. In a proof of concept with Apptronik and Mercedes-Benz, Apollo 2 sorted automotive parts into trays on production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The robot achieved a 96% success rate while correcting gaps in the workflow.
DeepMind also says new inference-time guidance improved safe task completion by 33% around obstacles and 43% near people. The company is releasing ASIMOV-Jailbreak, an open-source benchmark for testing unsafe robotics requests. Gemini Robotics 2.5 is being rolled out to robotics partners, with Gemini 4 Argon used to accelerate simulation and development.
Written by Sakhanda Wire AI from reporting by DeepMind. Facts, quotes and full context are in the original article below.