Robot navigation models crash in ways their makers don't measure
What happened
Researchers found that the usual way of testing robot navigation models hides serious flaws. These models frequently crash, get lost in similar-looking places, and fail when conditions change, even if they technically reach their destination.
Why it matters
For years, robot developers have relied on simple "success rates" to show their navigation models work. This paper shows those numbers are misleading. It becomes harder for companies to claim their autonomous robots are safe for complex real-world environments without more rigorous testing.
The signal
Watch whether robot developers and industry groups start adopting these new evaluation methods and datasets, or if they continue to rely on simpler metrics.