Gemini Robotics 2.5 and open safety benchmark from DeepMind

What happened

Google DeepMind has unveiled Gemini Robotics 2.5, a new model for robots, alongside an open safety benchmark called ASIMOV-jailbreak. According to reporting by Ars Technica, the model is up to twice as successful as its predecessor at some tasks.

The release pairs a capability update with a publicly available safety test. The benchmark's name points to the problem it targets: attempts to get an AI system to bypass its safety constraints. The available reporting does not set out the full task list, the evaluation method or the size of the gains beyond that headline figure, so those details should be checked against DeepMind's own documentation.

Why it matters

For organisations weighing up physical AI, the news signals two things moving together. Robots are getting more capable at specific tasks, and the tools to test whether they can be pushed into unsafe behaviour are being shared openly rather than kept in-house. Both affect how quickly robotics can move from pilots into customer-facing and operational settings such as logistics, retail and field services.

An open benchmark also gives buyers and service teams a common yardstick. When vendors report results against the same public tests, comparing options and setting acceptance criteria becomes more practical, and safety stops being a claim each supplier makes on its own terms.

By the numbers

• Up to 2x the task success rate of the previous model on some tasks, according to Ars Technica's summary of DeepMind's announcement.

The Renascence take

Most coverage will focus on the "twice as successful" headline. The more useful question for experience leaders is which tasks improved, and whether those are the ones that shape the customer's moment of truth.

A robot that succeeds twice as often on a narrow task still fails in ways customers remember. Behavioural economics tells us that a single visible failure weighs more heavily than many quiet successes, so the real measure is recovery design: how the system signals uncertainty, hands over to a human and makes the next step obvious. Operators should ask vendors for failure-mode data and recovery journeys, not just benchmark scores. Treat an open safety benchmark as the start of a service-design conversation, not a procurement checkbox.