预测性 Foot IK：让脚步先于地形，而不是踩空后再修正Predictive Foot IK: Anticipating Terrain Instead of Correcting After Contact
大多数 Foot IK 教程做的是被动纠正：每帧向下 trace，把脚贴到地面。平地没问题，但在楼梯、斜坡、碎石上会出现两类瑕疵 —— 脚在半空短暂穿模，或贴地时机不对导致「滑步」。
问题：纠正是「滞后」的
被动方案的本质是「先播动画，再事后拉回」。脚已经按平地假设摆到了错误位置，我们才在后处理里强行拽回地面 —— 视觉上就是一帧抖动或滑动。
思路：把「落地时机」变成可预测的数据
FootPlantAlpha_L / _R：0→1 表示这只脚从摆动进入「计划落地」的程度；FootState_L / _R：离散标记 planted / swinging。
float alpha = GetCurveValue("FootPlantAlpha_L");
FVector predicted = PredictFootTarget(LeftFootBone, alpha);
SetIKGoal(LeftFootGoal, predicted, alpha);骨盆偏移别忘了
两只脚落在不同高度时，要把骨盆 (pelvis) 向下偏移到较低那只脚的水平，否则腿会被拉直穿模。
手 K 曲线「贵」，但换来的是确定性：落地时机不再依赖运行时的猜测。
Most foot-IK tutorials do passive correction: trace down every frame and snap the foot to the ground. Fine on flat ground — but on stairs, slopes, and rubble you get two artifacts: the foot briefly clips mid-air, or it plants at the wrong moment and "slides".
The problem: correction lags
The passive approach is "play the animation, then yank it back". The foot has already moved to a wrong position assuming flat ground, and only then do we drag it to the surface — a frame of jitter or sliding.
The idea: turn "plant timing" into predictable data
FootPlantAlpha_L / _R: 0→1 for how far a foot has moved from swing into "planned plant";FootState_L / _R: discrete markers for planted / swinging.
float alpha = GetCurveValue("FootPlantAlpha_L");
FVector predicted = PredictFootTarget(LeftFootBone, alpha);
SetIKGoal(LeftFootGoal, predicted, alpha);Don't forget the pelvis offset
When the two feet land at different heights, lower the pelvis toward the lower foot, or the leg gets straightened and clips.
Hand-authored curves are "expensive", but they buy determinism: plant timing no longer relies on a runtime guess.