Robots Still Can't Tell If They Did a Good Job

Every robot that uses modern AI — the kind that can see a room, understand a spoken instruction, and move around — learns the same way a student does: by getting graded. Someone or something has to watch what the robot does and say "good job" or "bad job," thousands of times, during training and on the job. This paper surveys roughly 150 different ways of doing that grading, called "verifiers." They range from a human watching a video, to simple rules like "did it bump into anything?", to another AI model acting as a judge.

The team's big finding is what they call "no free checker." Think of it as a trade-off. Quick, cheap, constant feedback — like a motion sensor that instantly reports "no crash" — is easy to produce but easy to fool. A robot rewarded only for not crashing might learn to sit still forever and do nothing. Slow, expensive feedback — a person carefully reviewing footage, or strict mathematical rules — is far more trustworthy, but you can't get it every second, and it costs real money. In short: the more you get, the less you can trust it.

That trade-off has a name in AI: reward hacking. It's like a student who stuffs in filler words to hit an essay length instead of actually answering the question. For delivery robots, warehouse machines, self-driving cars, and the humanoid robots companies keep promising, this is the difference between a machine that looks safe in a demo and one that behaves safely in your neighborhood. Safety filters and real-time monitors catch problems as they happen — but they tend to be simple, and simple is gameable.

The paper closes with nine metrics that would make a company's claims about its robot checkers verifiable, plus a map of the gaps still to be filled. Until those checkers get better, expect slower, costlier progress toward robots you'd actually trust in your home or on your street.