Google’s AI research lab DeepMind has introduced Gemini Robotics 2.5, an updated vision-language-action (VLA) model designed to improve robotic dexterity, reliability, and physical safety in unstructured, real-world environments. Rolling out to robotics partners, the model works across various hardware platforms and was developed with support from Gemini 4 Argon, which acted as a simulation engine to generate virtual training environments and analyze task failures.
In the announcement they shared that in the evaluations covering more than 200 distinct tasks, such as using tongs to serve food, tying ribbons, and pipetting liquids, Gemini Robotics 2.5 achieved a 30% to 100% higher success rate and completed actions 20% to 30% faster than previous baselines. In a live proof-of-concept conducted with Apptronik (Automation company) and the German automotive brand Mercedes-Benz, at production plants in Hungary and Germany, Apptronik’s Apollo 2 humanoid robot used the model alongside Gemini Robotics ER to autonomously sort and pack automotive parts with a 96% success rate.
The release also introduces updates to physical safety and adversarial robustness. Using real-time inference-time reasoning guidance to navigate hazards, the system increased safe task completion by 33% around static obstacles and by 43% near people. Alongside the model, Google DeepMind released ASIMOV-Jailbreak, an open-source safety benchmark testing resistance to disguised unsafe instructions, on which Gemini 4 Argon recorded a 99.5% resistance rate.
