Gemini robotics 2.5 boosts robot dexterity and real-world adaptability

The introduction of Gemini Robotics 2.5 marks a significant advancement in vision-language-action (VLA) models, enhancing both the scope and speed of skills robots can learn, thereby improving dexterity and reliability for unstructured environments. Evaluations across more than 200 distinct tasks, such as tying a bow or serving food with tongs, demonstrated substantial gains; Gemini Robotics 2.5 achieved success rates 30% to 100% higher and completed tasks 20% to 30% faster than prior benchmarks. The model was tested on platforms like Apptronik’s Apollo 2 humanoid robot and the Franka Duo. In a proof-of-concept with Apptronik and Mercedes-Benz, Apollo 2 autonomously performed automotive kitting on production lines in Kecskemét, Hungary, and Sindelfingen, Germany, achieving a 96% success rate. Furthermore, safety features were advanced, with inference-time reasoning guidance boosting safe task completion around obstacles by 33% and near people by 43. Gemini 4 Argon also contributed by acting as a simulation engine, accelerating development and achieving 99.5% jailbreak resistance against the new ASIMOV-Jailbreak benchmark.