Robot Tried to Improve Itself 123 Times — Still Couldn't Stock a Fridge

A researcher at the National University of Singapore built an AI system that improves robots by itself — no human writing robot code. It watches a robot fail, figures out what skill is missing, writes or installs a fix, tests it in simulation, and repeats. He ran it for 123 rounds on ordinary household chores. The target job was simple to describe: put condiments on the top shelf of a fridge.

The good news: the AI could spot its own blind spots. When targets kept being out of view, it asked for a camera-viewing model, debugged it, and deployed a working 'look around' skill. The bad news: the improvements never added up. Each change passed its test, yet the real task never worked once.

The AI itself was rarely the problem. Three things around it were. First, today's vision tools don't understand relationships: software like SAM 3 finds 'a shelf,' but not 'the top shelf,' so the AI piled on more and more geometry rules that never clicked. Second, long tasks fail at the very first step, so all the evidence and fixes pile up there, while later steps are barely tested or improved. Third, the testing setup decides what gets learned. The AI optimized exactly what the evaluator measured — mistakes included — so weak tests and misleading memory turned busy work into standing still.

What does this mean for you? Claims that robots will teach themselves to cook, clean, or stock shelves are running ahead of reality. The bottleneck isn't ambition, it's measurement and perception: knowing what 'good' looks like and what the robot is actually seeing. The same lesson applies to any workplace automation. If you measure the wrong thing, you get a system that looks like it's improving while quietly missing the point — and that costs real time and money.