We noticed that the transforms return from tf/tf2 occasionally contain NaN values in the translation part. Some debugging revealed that this is caused by the interpolation logic of the time cache. The bug is hidden here:

The equality check compares the timestamps based on sec/nsec integers, however the ratio converts them to double. In our case the two timestamps are only a few nanoseconds apart, and the resolution after the toSec() conversion is not enough. Thus there is a division by zero and the interpolation returns meaningless results due to ratio being NaN.

I will open a PR with the most obvious fix: Checking that the toSec() difference is non-zero instead of the stamp equality check.