Motion Matching

September 9, 2024About 3 min

Motion Matching

Motion Matching is a simple idea, that helps us reason about movement description and control.

It’s also a new type of animation system, with three advantages:

High quality
Controllable responsiveness
Minimal manual work

播放动画

一段播放动画的代码：

// start
if (!walking && wantToWalk)
{
    PlayAnim(StartAnim);
    walking = true;
}

// walk loop
if (IsPlaying(StartAnim) && IsAtEndOfAnim())
{
    PlayAnim(WalkLoopAnim);
}

// stop
if (walking && !wantToWalk)
{
    PlayAnim(StopAnim);
    walking = false;
}

同样可以使用状态机，可视化：

但如果动画过多就会出现下图：

当然还有混合树的情况播放动画：

if (speed > 3.0f)
{
    PlayAnim(RunAnim);
}
else if (speed > 0.0f)
{
    PlayAnim(WalkAnim);
}
else
{
    PlayAnim(IdleAnim);
}

同样也可以使用可视化工具来处理：

多个参数控制动画

如果有多个数据都会影响着动画的播放，就是需要配表来控制标准选择动画，然后实时传递值给动画控制器。这个逻辑就是动画蓝图的功能。

改变固定动画播放路径

Motion Graphs

选择下一个动画

Reinforcement Learning Based Character Locomotion in Hitman: Absolution

Motion Fields for Interactive Character Animation

Motion Fields 性能消耗非常大，这里只使用了跳转到其他动画的任何一帧。

选择正确的开始和结束时间

姿势匹配
速度匹配
精确的末端位置匹配

原理简介

核心逻辑：每一帧都去寻找最合适的动画，然后跳转播放。

Mocap

先做动作捕捉

Code

下面是每帧计算匹配的简单代码：

int m_CurrentAnimIndex;
int m_CurrentAnimTime;

void AmoUpdate(Goal goal, float dt)
{
    m_CurrentAnimTime += dt;
    Pose currentPose = EValuateLerpedPoseFromData(m_CurrentAnimIndex, m_CurrentAnimTime);

    float bestCost = 1000000;
    Pose bestPose;

    // loop on all mocap
    for (int i = 0; i < m_Poses.Size(); i++)
    {
        Pose candidatePose = m_Poses[i];

        // every candidate jumping point has a cost
        float thisCost = ComputeCost(currentPose, candidatePose, goal);
        if (thisCost < bestCost)
        {
            // remember the best candidate
            bestCost = thisCost;
            bestPose = candidatePose;
        }
    }

    bool theWinnerIsAtTheSameLocation = m_CurrentAnimIndex == bestPose.m_AnimIndex && fabs(m_CurrentAnimTime - bestPose.m_AnimTime) < 0.2f;

    if(!theWinnerIsAtTheSmaeLocation)
    {
        // blend to the winning location
        m_CurrentAnimIndex = bestPose.m_AnimIndex;
        m_CurrentAnimTime = bestPose.m_AnimTime;
        PlayAnimStartingAtTime(m_CurrentAnimIndex, m_CurrentAnimTime, 0.25f);
    }
}

Trick 1: Posematch only a few bones

Local velocity
Feet positions
Feet velocities
Weapon position

Trick 2: Just check where a piece of animation brings you if you play it

动画的目标，部分代码：

class TrajectoryPoint
{
    Vector3 m_Position;
    float m_Sight;
    float m_TimeDelay;
};

// desired goal, sent by gameplay each frame
class Goal
{
    Array<TrajectoryPoint> m_DesiredTrajectory;
    Stance m_DesiredStance;

    // ...
};

计算调整到新的姿势的花销的代码：

float ComputeCost(Pose currentPose, Pose candidatePose, Goal goal)
{
    float cost = 0.0f;

    // how much the candidate jumping position matches the current situation
    cost += ComputeCurrentCost(currentPose, candidatePose);

    // this is our responsivity slider
    static float reponsivity = 1.0f;

    // how much the candidate piece of motion matches the desired trajectory
    cost += responsivity * ComputeFutureCost(candidatePose, goal);

    return cost;
}