New Software Spots Robot Mistakes Before They Happen

Most robot safety software works like a referee who only blows the whistle after the foul. It watches what a machine actually does, and when something goes wrong, it files a report. That's fine for a video game. It's much less fine for a robot crawling through a nuclear waste store, where one bad move can be expensive, dangerous, or impossible to undo. Researchers at the University of Manchester and the University of York have published a new approach that tries to change the timing of that warning.

Their system, called Predictive Varanus, works in two stages. First, it checks whether the robot is still following its own rules — the written plan that says what it may do and in what order. If the robot drifts outside that plan, the system flags it immediately. Then, for robots that are still behaving correctly so far, a second layer looks ahead at the possible next moves and asks a sharper question: is this heading toward a guaranteed problem, or is it still fine? That means the software can raise an alarm several steps early, while there is still time to stop or redirect the machine.

The clever part is that it reuses rules engineers already write during the design phase, rather than asking them to build a second, separate safety system from scratch. In their demonstration, the authors applied it to a robotic rover inspecting a nuclear storage facility — a scenario where a mistake isn't just a bug, it's a hazard.

The catch: this is a research paper, not a product. There's no release date, no price, and no evidence yet from real factories or hospitals. It also only works for robots whose rules can be written down precisely in advance — messier, unpredictable environments remain a much harder problem.