Google DeepMind has launched Gemini Robotics 2.5, its most capable vision-language-action model for robots. The model improves dexterity, reliability, and safety, allowing robots to learn a wider range of skills faster. In evaluations across more than 200 tasks, including handling tongs, tying bows, and pipetting, it achieved 30 to 100 percent higher success rates and completed tasks 20 to 30 percent faster than previous versions. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots used the model to pick automotive parts from bulk bins and arrange them into trays, reaching a 96 percent success rate at factories in Hungary and Germany. The model can detect hazards and people nearby and adjust movements in real time, improving safe task completion by 33 percent around obstacles and 43 percent around people. Google also released the open-source ASIMOV-Jailbreak benchmark to test resistance to unsafe instructions.
Google DeepMind has launched Gemini Robotics 2.5, its most capable vision-language-action model for robots. The model improves dexterity, reliability, and safety, allowing robots to learn a wider range of skills faster. In evaluations across more than 200 tasks, including handling tongs, tying bows, and pipetting, it achieved 30 to 100 percent higher success rates and completed tasks 20 to 30 percent faster than previous versions. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 humanoid robots used the model to pick automotive parts from bulk bins and arrange them into trays, reaching a 96 percent success rate at factories in Hungary and Germany. The model can detect hazards and people nearby and adjust movements in real time, improving safe task completion by 33 percent around obstacles and 43 percent around people. Google also released the open-source ASIMOV-Jailbreak benchmark to test resistance to unsafe instructions.