· 24–30 August 2026
Better calibration, no safer robot
Research 1's confirmatory analysis ran against a plan frozen before the data was touched. Calibration improved and clean-route efficiency held, and both preregistered safety hypotheses failed anyway. Freezing the analysis is what made that an answer rather than an embarrassment.
This week
The confirmatory analysis for Research 1 was executed, the release inventory was verified against a clean clone, and the core dataset was deposited.
The question was whether converting calibrated semantic perception uncertainty into navigation cost makes a Nav2 robot safer under controlled distribution shift. The answer is no, or at least not demonstrably, and the useful part is that the analysis plan said what would count as yes before anyone looked.
Results
Calibration improved substantially. Clean-route efficiency was preserved, which matters because the cheap way to reduce collisions is to make the robot timid and slow, and that would have shown up here.
The preregistered collision-reduction and severity-interaction hypotheses did not pass. Better component calibration did not establish safer navigation. Those are two separate claims and the study only supports the first.
Underneath the aggregate, the paired comparison holds 29 worsenings against 14 improvements — a distribution that a headline mean would have flattened into something more flattering than the evidence supports.
Failures
Several fault families produced zero events, which drives quasi-complete separation and makes the model estimates unstable rather than merely imprecise. It is named as that in the write-up instead of being reported as a very large effect, which is what an unexamined fit would have produced.
A related trap: sampling convergence is not calibration. A chain that has converged tells you the sampler behaved, not that the resulting uncertainties are trustworthy, and the two get conflated constantly.
Open questions
Whether the null is a property of this pipeline or of the idea. The costmap path fuses semantic cost softly and generically, so a negative result here is consistent both with "calibrated uncertainty does not help navigation" and with "this particular fusion wastes it". Distinguishing those needs a different experiment, not a reanalysis of this one.
Next
Freeze the taxonomy, reconcile the publication surfaces with what the analysis actually licenses, then put the whole thing in front of an independent reader.