Same task: get past the obstacle. "Go left" and "go right" are both valid — a genuinely bimodal future. Drag toward point estimate (or naive time-averaging) and the two modes collapse into their mean: a straight line through the obstacle — the most dangerous action of all, one nobody actually proposed. Keep it a distribution and both safe detours survive. A flow/diffusion policy already outputs a distribution; the lesson is don't average it down to a point.