Robots Were Being Punished for Acting Safely — That Bug Is Now Fixed
A robot saying 'I'll avoid the lawn' was graded as if it drove on it.
An open-source robot safety project called QERRA-THRIVE released a fix this week for a surprisingly human problem: it couldn't understand the word 'not.' The system works like a referee for robots. Before a robot does anything, it generates a list of possible plans, and this software grades each one across 12 categories — things like human safety, animal awareness, honesty, and care for the environment. The highest-scoring plan wins.
The bug was embarrassing. The grader was penalizing robots that mentioned a hazard in order to say they were avoiding it. A plan that said 'Avoiding driving over the lawn to reach the drop zone' was docked points for a lawn violation. A robot saying 'I am not qualified to service this high-voltage assembly, requesting a supervisor' was penalized for being unqualified. Worst of all, one robot said 'Will not maintain current pace, but will slow down around personnel' — and the software read the 'will not' as refusing to slow down, skipping right past the comma. Careful robots lost 0.15 to 0.30 points just for being polite, which meant careless plans that said nothing at all were winning.
The fix is straightforward once you see it. The software now looks backward about 40 characters before flagging a hazard, checking whether the robot was actually describing avoidance. It also stops its pattern-matching at commas, periods, and semicolons, so a 'not' in one sentence can't leak into the next one. The developer tested it against deliberately tricky 'negative twin' examples and all 12 categories now pass, with the safe, thoughtful plans winning by healthy margins.
Why should you care? Self-driving cars, delivery robots, and warehouse machines all face this exact problem. An AI that misreads 'do not enter' as 'enter' is dangerous. Getting the word 'no' right, reliably, is one of the quiet foundations of letting robots operate near people at all.
- The software grades robot plans on 12 categories, including human safety, honesty, and environmental care, and picks the best one to execute.
- The bug docked safety-conscious plans 0.15 to 0.30 points — enough to let careless plans that mentioned no hazards win instead.
- The fix teaches it to look at the words right before a hazard and stop reading at commas, so 'avoid' and 'will not' mean what they say.
Why It Matters
Robots that explain their caution should be trusted more, not less — this fix makes that possible.