Google DeepMind announces Gemini Robotics 2.5 and ASIMOV-Jailbreak

In short: Google DeepMind unveiled Gemini Robotics 2.5, a new version of its robot AI, and published an open safety test called ASIMOV-Jailbreak.

What happened

Google DeepMind says Gemini Robotics 2.5 helps robots do hands-on tasks more reliably, especially work that needs careful finger and hand control. In tests across more than 200 activities, DeepMind reported a 30 to 100 percent higher success rate than the previous version, and it was 20 to 30 percent faster. The company also said the new model is up to twice as successful at some tasks.

DeepMind showed examples like serving food with tongs, using a lab pipette (a small tool for moving tiny amounts of liquid), and tying a bow with a five-fingered humanoid robot. Tying ribbon is hard for robots because soft items bend and change shape, unlike a rigid tool or a box.

Google worked with partners including Apptronik and Mercedes-Benz to try the system in real factories, not just in a lab. DeepMind said Gemini-powered Apollo 2 robots did parts-kitting work in a few European Mercedes-Benz factories and reached a 96 percent success rate. DeepMind also noted that getting from 96 percent to near-perfect performance is still difficult.

To address safety, DeepMind released ASIMOV-Jailbreak, an open source set of 1,500 written prompts meant to trick a robot into doing something unsafe. Think of it like a driving test that includes tricky situations on purpose. The dataset is available on Hugging Face.

Why it matters

If general-purpose robots are going to work near people, they need to be both capable and predictable. Open safety tests like ASIMOV-Jailbreak give researchers and companies a shared way to check whether robots will refuse unsafe instructions.

Source: Arstechnica