DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"

A five‑fingered humanoid robot, Apptronik’s Apollo 2, just tied a bow with Gemini 2.5, a feat that shows the system can handle deformable objects that have long stumped robots. The trick hinges on Gemini Embodied Reasoning, a vision‑language model that watches live video and plans each move before the robot acts.

Gemini Robotics 2.5, the action engine, then turns that plan into precise actuator commands. DeepMind tested the new model on more than 200 activities, from serving food with tongs to pipetting in a lab. The upgrade delivers a noticeable jump in dexterity, especially for tasks that demand fine motor control.

Yet the system still relies on a multi‑layer pipeline, so any lag in perception can ripple into a misstep. In a pilot at Mercedes‑Benz factories, Gemini robots completed real‑world jobs, proving they can operate in a production line environment. The robots followed human‑provided instructions and used live video to navigate the workspace, showing that the vision‑language loop works beyond the lab.

Google is preparing an open‑source safety benchmark that will let developers test how well robots can coexist with people.

The benchmark could set a new standard for safe human‑robot collaboration, but it also highlights that the technology is still experimental and needs rigorous testing before widespread deployment.