Google DeepMind has launched Gemini Robotics 2.5, an upgraded vision-language-action (VLA) model that drives robotic actuators, paired with the Gemini Embodied Reasoning vision-language planner. Evaluated across over 200 tasks—including dexterous manipulation such as serving food with tongs, laboratory pipetting, and tying a bow with Apptronik’s Apollo 2 humanoid’s five-fingered hands—the new version achieves a 30–100% higher success rate and operates 20–30% faster than its predecessor. In real-world validation at Mercedes-Benz European factories, Gemini-powered Apollo 2 robots kitted automotive parts with a 96% success rate. DeepMind director Kanishka Rao acknowledges that closing the remaining 4% gap to human-level reliability represents a steep challenge. Concurrently, DeepMind released ASIMOV-jailbreak, an open-source safety benchmark designed to evaluate and mitigate jailbreak vulnerabilities in embodied AI systems, supporting safer deployment of robots operating alongside humans.