DeepMind Unveils Gemini Robotics 2.5 and an Open Robot Safety Benchmark

Google DeepMind has introduced the 2.5 version of Gemini Robotics, the vision-language-action layer that turns instructions into actual robot movement, like driving an actuator. It sits under a second piece called Gemini Embodied Reasoning, a vision-language model that reads your command and uses live video to plan how the robot gets the job done. Google says the result is a robotic AI that is closing in on human-level performance at real-world tasks.

The headline upgrade is dexterity. DeepMind evaluated Gemini Robotics 2.5 across more than 200 activities, including serving food with tongs and handling a pipette the way a lab tech would. The team also showed Apptronik's Apollo 2 humanoid, running the 2.5 model, tying a bow with its five-fingered hands. That is a genuinely hard trick, because deformable objects like ribbon have long been a nightmare for robots to manipulate.

This is not just demo-reel material. According to Ars Technica, some Gemini-powered robots have already done real work inside Mercedes-Benz factories. That puts DeepMind's physical AI platform past the lab stage and onto production floors, where mistakes cost money and time.

DeepMind is also rolling out an open source safety benchmark, reported as ASIMOV-jailbreak, aimed at the day robots operate alongside humans. The company says the release is about getting ready for that shared workspace rather than testing machines in isolation.