Detecting estimation failure

[hidmic]

There are many conditions in which an observer may deviate from the true yet unobservable state it is trying to estimate. A bad measurement, a changing environment, even intrinsic limitations of the observer itself (limited sample size, local minima) may lead to failure. Particle filters are not free from this, and neither is Beluga. In fact, this has been a common concern with good ol' AMCL for a very long time, so much so that localization jump detectors have been developed by many over time.

Can we detect failure? I´d naively suggest statistical hypothesis testing to look for outliers in reported measurements and state estimates. After all, PFs have complete probability distributions to work with. Is that enough?

[hidmic]

@pepisg any insights on how you guys have been detecting failures would be valuable.