Google DeepMind has revealed Gemini Robotics 2.5, an update to its robot-control AI that the company says is getting close to matching humans on real-world tasks.
The system works in two layers. Gemini Embodied Reasoning, a vision-language model, takes an instruction and uses live video to plan how a robot should do the job. Gemini Robotics, now at version 2.5, is a vision-language-action model, meaning it turns that plan into physical movements such as driving an actuator (the motor-powered part that moves a joint or gripper).
DeepMind says the new version is much better at tasks needing fine dexterity. It was evaluated across more than 200 activities, including serving food with tongs and using a pipette. Google also showed Apptronik's Apollo 2 humanoid, running the model, tying a bow with its five-fingered hands. Ribbon is a deformable object, and those have long been hard for robots to handle because the material changes shape as it is touched.
Some Gemini robots also did actual work in Mercedes-Benz factories. Google has not said how many robots were involved, which tasks they handled or for how long, and those details decide whether this is a pilot or the start of something larger.
The companion release is an open source safety benchmark called ASIMOV-jailbreak. DeepMind framed it as preparation for robots operating alongside humans. The name suggests a test of whether a robot can be talked out of its safety rules, a problem chatbots already know well. A chatbot that is jailbroken produces bad text; a robot that is jailbroken moves in a room with people in it.
The pairing is the notable part. Dexterity demos are easy to find, but a published safety test is a harder commitment to make. The open question is how Gemini Robotics 2.5 scores on it, and whether rivals building robot models run the same test.