ETH官方钱包

前往
大廳
主題

狼 追逐 v2 因為夠帥氣,所以從今天開始你的名字是魔狼妖吼!

夏洛爾 | 2022-12-12 22:27:05 | 巴幣 0 | 人氣 171


Wolf Run V2

實驗目標:
1.進入靜立狀態後,進入追逐狀態,在追逐狀態下,要能持續跑至接近目標的距離內
2.尺寸介於1-5倍

實驗設計:
1.任何弱點觸地皆失敗 (尾巴和四個小腿並非是弱點)
2.非弱點肢體
if(wolfBodies[i].damageCoef > 0f){clampReward += -0.01f * wolfBodies[i].damageCoef;}
3.
//Set: judge.endEpisode = false//Set: useClampReward = true//Set: SharpingBuffer Len=250 Th=-0.4if(weaknessOnGround){if(inferenceMode){brainMode = BrainMode.GetUp;SetModel("WolfGetUp", getUpBrain);behaviorParameters.BehaviorType = BehaviorType.InferenceOnly;}else{AddReward(-1f);judge.outLife++;judge.Reset();return;}}else if(wolfRoot.localPosition.y < -10f){if(inferenceMode){brainMode = BrainMode.GetUp;SetModel("WolfGetUp", getUpBrain);behaviorParameters.BehaviorType = BehaviorType.InferenceOnly;}else{AddReward(-1f);judge.outY++;judge.Reset();return;}}else{targetSmoothPosition = targetPositionBuffer.GetSmoothVal();headDir = targetSmoothPosition - stageBase.InverseTransformPoint(wolfHeadRb.position);rootDir = targetSmoothPosition - stageBase.InverseTransformPoint(wolfRootRb.position);flatTargetVelocity = rootDir;flatTargetVelocity.y = 0f;targetDistance = flatTargetVelocity.magnitude;Vector3 forwardDir = flatTargetVelocity.normalized;Vector3 flatLeftDir = Vector3.Cross(flatTargetVelocity, Vector3.up);lookAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfHead.right * -1f, headDir));//SideUpupAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfHead.forward, flatLeftDir));aimVelocity = flatTargetVelocity.normalized;aimVelocity.y = 0.2f;//SideUpspineUpAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfSpine.forward, flatLeftDir));rootUpAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfRoot.right*-1f, flatLeftDir));leftThighAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfLeftThigh.forward * -1f, flatLeftDir));rightThighAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfRightThigh.forward * -1f, flatLeftDir));//For Sync runVector3 leftThighUpDir = Vector3.ProjectOnPlane(wolfLeftThigh.right, flatLeftDir);Vector3 rightThighUpDir = Vector3.ProjectOnPlane(wolfRightThigh.right, flatLeftDir);float thighUpAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(leftThighUpDir, rightThighUpDir));leftUpperArmAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfLeftUpperArm.forward * -1f, flatLeftDir));rightUpperArmAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfRightUpperArm.forward * -1f, flatLeftDir));//For Sync runVector3 leftUpperArmUpDir = Vector3.ProjectOnPlane(wolfLeftUpperArm.right, flatLeftDir);Vector3 rightUpperArmUpDir = Vector3.ProjectOnPlane(wolfRightUpperArm.right, flatLeftDir);float upperArmUpAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(leftUpperArmUpDir, rightUpperArmUpDir));tailAngle = Mathf.InverseLerp(180f, 0f, Vector3.Angle(wolfTail.right, flatTargetVelocity));avgVelocity = velocityBuffer.GetSmoothVal();velocityAngle = Vector3.Angle(avgVelocity, aimVelocity);velocityAngleCoef = Mathf.InverseLerp(180f, 0f, velocityAngle);flatVelocity = avgVelocity;flatVelocity.y = 0f;flatVelocityManitude = flatVelocity.magnitude;float sizeScale = Mathf.Lerp(1f, 2.5f, currentSize/5f);velocityCoef = Mathf.InverseLerp(0f, 15f*sizeScale, Vector3.Project(avgVelocity, aimVelocity).magnitude );flatVelocityAngle = Vector3.Angle(flatVelocity, flatTargetVelocity);if(!inferenceMode){if(targetDistance > nearModeRange){if(Time.fixedTime - landingMoment > landingBufferTime){bool outSpeed = flatVelocityManitude < Mathf.Lerp(0f, 7f*sizeScale, (Time.fixedTime - landingMoment - landingBufferTime)/4f);bool outDirection = flatVelocityAngle > Mathf.Lerp(180f, 30f, (Time.fixedTime - landingMoment - landingBufferTime)/5f);float motionLimit = Mathf.Lerp(0f, 0.8f, (Time.fixedTime - landingMoment - landingBufferTime)/3f);float motionLimit2 = Mathf.Lerp(0.3f, 0.7f, (Time.fixedTime - landingMoment - landingBufferTime)/3f);float sharpingResetVal = Mathf.Lerp(0f, sharpingResetThreshould, (Time.fixedTime - landingMoment - landingBufferTime - 2f)/5f);bool outMotion = lookAngle < motionLimit2 || upAngle < motionLimit2 || leftThighAngle < motionLimit2 || rightThighAngle < motionLimit2 || spineUpAngle < motionLimit || rootUpAngle < motionLimit || thighUpAngle < motionLimit2 || upperArmUpAngle < motionLimit2 || leftUpperArmAngle < motionLimit2 || rightUpperArmAngle < motionLimit2;if( outSpeed || outDirection || outMotion){// AddReward(-1f);if(outSpeed){#if UNITY_EDITORDebug.Log("outSpeed");#endifclampReward += -0.05f;judge.outSpeed++;}if(outDirection){#if UNITY_EDITORDebug.Log("outDirection");#endifclampReward += -0.05f;judge.outDirection++;}if(outMotion){#if UNITY_EDITORDebug.Log("outMotion");#endifclampReward += -0.05f;judge.outMotion++;}sharpingBuffer.PushVal(-1f);// judge.Reset();// return;}else{sharpingBuffer.PushVal(0f);}#if UNITY_EDITORsharpingVal = sharpingBuffer.GetSmoothVal();#endifif( sharpingBuffer.GetSmoothVal() < sharpingResetVal){AddReward(-1f);judge.Reset();return;}}lastReward = (velocityAngleCoef + velocityCoef) * 0.02f + (lookAngle+upAngle) * 0.01f + (leftThighAngle+rightThighAngle+leftUpperArmAngle+rightUpperArmAngle) * 0.005f+ (spineUpAngle+rootUpAngle) * 0.005f+ (tailAngle) * 0.005f+ (thighUpAngle + upperArmUpAngle) * 0.005f+ (1f - exertionRatio) * 0.005f;if(useClampReward){lastReward = lastReward+clampReward;if(lastReward < -0.05f) lastReward = -0.05f;}totalReward += lastReward;AddReward( lastReward );}// else if(targetDistance > 1.5f)else{// AddReward(1f);judge.survived++;judge.Reset();return;}}}

//大致來說,
--1.獎勵視線,並使用Force Sharping
--2.獎勵投影至"跑動推薦向量"的速度和角度,並使用Force Sharping
--3.獎勵四個大腿的Side Look,並使用Force Sharping
--4.獎勵尾巴符合指定角度
--5.獎勵減少動作變化
--6.獎勵雙手和雙足要同步奔跑
--7.Motion相關的Force Sharping非從0開始

4.Force Sharping改為有容錯空間,但是容許值逆向Sharping
允許角色在5秒內發生總計2秒以內的失誤,希望藉此讓角色就算輕微失衡也能嘗試自行修正
但是容許值是逆向Sharping,會在開始Force Sharping後兩秒才逐步放寬標準

實驗時間:
Step: 5e7
Time Elapsed: 94401s (26.22hr)

實驗結果:
實驗結果為成功,但不理想

藉由強制雙手雙腳需要同步奔跑,狼有雙手雙腳同步奔跑了

但是問題有三

1.其實只有三隻腳
左手幾乎是沒有用上,只有偶爾才輔助

2.小尺寸奔跑能力較弱
目前看起來是因為大部分會因Out Of Speed被Force Sharping淘汰
這裡認為是非對稱問題,由於不希望體型大的狼跑太快,所以體型越大速度要求越低
但看來反而讓同動作,體型小的狼無法抵達目標速度,體型大才可以
但因為本實驗沒有設定極限時間,所以又發生訓練量不均勻的問題

3.狼會翹起來
不知為何身體曲線抬很高
檢查後也不覺得狼有前後長短腳,所以感覺是因為進入Gait
初期亂加速就很容易抬起來,然後因為也沒差,就所幸一直抬著

另外感覺左右搖擺的情況還是很嚴重

因此下個實驗為狼追逐
1.引導身體要盡量平行地面
2.速度要求正比尺寸
3.提高ForceSharping的要求,尤其進入階段

創作回應

更多創作