unity-tech-cn
/
ml-agents
45
14
派生 39
代码 合并请求 版本发布
浏览代码

more cleanup

/active-variablespeed
HH 4 年前
当前提交
40b5cb96
共有 1 个文件被更改，包括 40 次插入 和 188 次删除
  1. 228
      Project/Assets/ML-Agents/Examples/Walker/Scripts/WalkerAgent.cs

228
Project/Assets/ML-Agents/Examples/Walker/Scripts/WalkerAgent.cs
查看文件


public class WalkerAgent : Agent
{
[Header("Walk Speed")]
[Range(0, 15)]
public float walkingSpeed = 15; //The walking speed to try and achieve
[Header("Walk Speed")] [Range(0, 15)] public float walkingSpeed = 15; //The walking speed to try and achieve
[Header("Target To Walk Towards")]
public TargetController target; //Target the agent will walk towards.
[Header("Target To Walk Towards")] public TargetController target; //Target the agent will walk towards.
[Header("Body Parts")] public Transform hips;
public Transform chest;

JointDriveController m_JdController;
EnvironmentParameters m_ResetParams;
// private WalkGroup walkGroup;
// walkGroup = FindObjectOfType<WalkGroup>();
//Setup each body part
m_JdController = GetComponent<JointDriveController>();
m_JdController.SetupBodyPart(hips);

orientationCube.UpdateOrientation(hips, target.transform);
rewardManager.ResetEpisodeRewards();
walkingSpeed = randomizeWalkSpeedEachEpisode? Random.Range(0.0f, m_maxWalkingSpeed): walkingSpeed; //Random Walk Speed
walkingSpeed =
randomizeWalkSpeedEachEpisode ? Random.Range(0.0f, m_maxWalkingSpeed) : walkingSpeed; //Random Walk Speed
SetResetParameters();
}

/// </summary>
public override void CollectObservations(VectorSensor sensor)
{
// avgVelValue = GetVelocity();
// sensor.AddObservation(VelocityInverseLerp(cubeForward * walkingSpeed, avgVelValue));
// sensor.AddObservation(VelocityInverseLerp( cubeForward * walkGroup.walkingSpeed)); //
sensor.AddObservation(VelocityInverseLerp( cubeForward * walkingSpeed));
sensor.AddObservation(GetMatchingVelocityInverseLerp(cubeForward * walkingSpeed, GetAvgVelocity()));
sensor.AddObservation(walkingSpeed/m_maxWalkingSpeed);
// sensor.AddObservation(walkGroup.walkingSpeed/walkGroup.m_maxWalkingSpeed);
sensor.AddObservation(walkingSpeed / m_maxWalkingSpeed);
sensor.AddObservation(Quaternion.FromToRotation(hips.forward, cubeForward));
sensor.AddObservation(Quaternion.FromToRotation(head.forward, cubeForward));

UpdateRewards();
}
Vector3 GetVelocity()
{
//Returns the average velocity of all the rigidbodies
Vector3 GetAvgVelocity()
{
Vector3 avg = Vector3.zero;
// velSum += m_JdController.bodyPartsDict[hips].rb.velocity;
// velSum += m_JdController.bodyPartsDict[spine].rb.velocity;
// velSum += m_JdController.bodyPartsDict[chest].rb.velocity;
// velSum += m_JdController.bodyPartsDict[head].rb.velocity;
// avg = velSum/4;
Vector3 avgVel = Vector3.zero;
int counter = 0;
int numOfRB = 0;
counter++;
numOfRB++;
avg = velSum/counter;
return avg;
// velInverseLerpVal = VelocityInverseLerp(cubeForward * walkingSpeed, avgVelValue);
avgVel = velSum / numOfRB;
return avgVel;
public float velInverseLerpVal;
public float hipsVelMag;
public float lookAtTargetReward; //reward for looking at the target
public float matchSpeedReward; //reward for matching the desired walking speed.
public float headHeightOverFeetReward; //reward for standing up straight-ish
public float hurryUpReward = -1; //don't waste time
// public float headHeightOverFeetReward; //reward for standing up straight-ish
public float bpVelPenaltyThisStep = 0;
//This reward will approach 1 if it matches and approach zero as it deviates
// matchSpeedReward =
// Mathf.Exp(-0.1f * (cubeForward * walkingSpeed -
// m_JdController.bodyPartsDict[hips].rb.velocity).sqrMagnitude);
hipsVelMag = m_JdController.bodyPartsDict[hips].rb.velocity.magnitude;
// velInverseLerpVal =
// Mathf.InverseLerp(0, walkingSpeed, m_JdController.bodyPartsDict[hips].rb.velocity.magnitude);
//This reward will approach 1 if it matches perfectly and approach zero as it deviates
var matchSpeedReward = GetMatchingVelocityInverseLerp(cubeForward * walkingSpeed, GetAvgVelocity());
// var moveTowardsTargetReward = Vector3.Dot(cubeForward,
// Vector3.ClampMagnitude(m_JdController.bodyPartsDict[hips].rb.velocity, maximumWalkingSpeed));
// b. Rotation alignment with goal direction.
// lookAtTargetReward = Vector3.Dot(cubeForward, head.forward);
lookAtTargetReward = (Vector3.Dot(cubeForward, head.forward) + 1) * .5F;
// lookAtTargetReward =
// Mathf.Exp(-0.1f * (cubeForward * walkingSpeed -
// m_JdController.bodyPartsDict[hips].rb.velocity).sqrMagnitude);
// b. Rotation alignment with target direction.
//This reward will approach 1 if it faces the target direction perfectly and approach zero as it deviates
var lookAtTargetReward = (Vector3.Dot(cubeForward, head.forward) + 1) * .5F;
headHeightOverFeetReward =
Mathf.Clamp01(((head.position.y - footL.position.y) + (head.position.y - footR.position.y))/ 10); //Should normalize to ~1
// AddReward(
// +0.02f * moveTowardsTargetReward
// + 0.01f * lookAtTargetReward
// + 0.01f * headHeightOverFeetReward
// );
// rewardManager.UpdateReward("matchSpeed", matchSpeedReward);
// rewardManager.UpdateReward("lookAtTarget", lookAtTargetReward);
// rewardManager.UpdateReward("headHeightOverFeet", headHeightOverFeetReward);
// rewardManager.UpdateReward("hurryUp", hurryUpReward/MaxStep);
// //VELOCITY REWARDS
// bpVelPenaltyThisStep = 0;
// foreach (var item in m_JdController.bodyPartsList)
// {
// var velDelta = Mathf.Clamp(item.rb.velocity.magnitude - walkingSpeed, 0, 1);
// bpVelPenaltyThisStep += velDelta;
// }
// rewardManager.UpdateReward("bpVel", bpVelPenaltyThisStep);
// headHeightOverFeetReward =
// Mathf.Clamp01(((head.position.y - footL.position.y) + (head.position.y - footR.position.y))/ 10); //Should normalize to ~1
// Vector3 velSum = Vector3.zero;
// avgVelValue = Vector3.zero;
// velSum += m_JdController.bodyPartsDict[hips].rb.velocity;
// velSum += m_JdController.bodyPartsDict[spine].rb.velocity;
// velSum += m_JdController.bodyPartsDict[chest].rb.velocity;
// velSum += m_JdController.bodyPartsDict[head].rb.velocity;
// avgVelValue = velSum/4;
// velInverseLerpVal = VelocityInverseLerp(cubeForward * walkingSpeed, avgVelValue);
velInverseLerpVal = VelocityInverseLerp(cubeForward * walkingSpeed);
rewardManager.rewardsDict["matchSpeed"].rewardThisStep = velInverseLerpVal;
rewardManager.rewardsDict["matchSpeed"].rewardThisStep = matchSpeedReward;
rewardManager.rewardsDict["headHeightOverFeet"].rewardThisStep = headHeightOverFeetReward;
// velInverseLerpVal = VelocityInverseLerp(cubeForward * walkGroup.walkingSpeed);
rewardManager.UpdateReward("productOfAllRewards", velInverseLerpVal * lookAtTargetReward);
// rewardManager.UpdateReward("productOfAllRewards", velInverseLerpVal * lookAtTargetReward * headHeightOverFeetReward);
// velInverseLerpVal = VelocityInverseLerp(Vector3.zero, cubeForward * walkingSpeed, avgVelValue);
// rewardManager.rewardsDict["headHeightOverFeet"].rewardThisStep = headHeightOverFeetReward;
rewardManager.UpdateReward("productOfAllRewards", matchSpeedReward * lookAtTargetReward);
}
//This reward will approach 1 if it matches and approach zero as it deviates
// velInverseLerpVal =
// Mathf.InverseLerp(0, walkingSpeed, avgVelValue.magnitude);
// rewardManager.UpdateReward("productOfAllRewards", velInverseLerpVal * lookAtTargetReward * headHeightOverFeetReward);
// matchSpeedReward =
// Mathf.Exp(-0.1f * (cubeForward * walkingSpeed -
// avgVelValue).sqrMagnitude);
// matchSpeedReward =
// Mathf.Exp(-0.01f * (cubeForward * walkingSpeed -
// avgVelValue).sqrMagnitude);
// rewardManager.UpdateReward("productOfAllRewards", matchSpeedReward * lookAtTargetReward * headHeightOverFeetReward);
// Vector3 velSum = Vector3.zero;
//
// int counter = 0;
// avgVelValue = Vector3.zero;
// foreach (var item in m_JdController.bodyPartsList)
// {
// counter++;
// velSum += item.rb.velocity;
// }
// avgVelValue = velSum/counter;
// //This reward will approach 1 if it matches and approach zero as it deviates
// matchSpeedReward =
// Mathf.Exp(-0.1f * (cubeForward * walkingSpeed -
// avgVelValue).sqrMagnitude);
// rewardManager.UpdateReward("productOfAllRewards", matchSpeedReward * lookAtTargetReward * headHeightOverFeetReward);
}
public Vector3 bodyVelocity;
//value of 0 means we are matching velocity perfectly
//value of 1 means we are not matching velocity
public float velDeltaDistance; //distance between the goal and actual vel
public float VelocityInverseLerp(Vector3 velocityGoal)
public float GetMatchingVelocityInverseLerp(Vector3 velocityGoal, Vector3 actualVelocity)
bodyVelocity = GetVelocity();
//distance between our actual velocity and goal velocity
var velDeltaMagnitude = Mathf.Clamp(Vector3.Distance(actualVelocity, velocityGoal), 0, walkingSpeed);
// velDeltaDistance = Vector3.Distance(avgVelValue, velocityGoal);
// velDeltaDistance = Vector3.Distance(avgVelValue, velocityGoal);
// velDeltaDistance = Mathf.Clamp(Vector3.Distance(bodyVelocity, velocityGoal), 0, walkGroup.walkingSpeed);
velDeltaDistance = Mathf.Clamp(Vector3.Distance(bodyVelocity, velocityGoal), 0, walkingSpeed);
// float percent = Mathf.InverseLerp(walkingSpeed, 0, velDeltaDistance);
// float percent = Mathf.InverseLerp(velocityGoal.magnitude, 0, velDeltaDistance);
// float percent = Mathf.Pow(1 - Mathf.Pow(velDeltaDistance/walkGroup.walkingSpeed, 2), 2);
float percent = Mathf.Pow(1 - Mathf.Pow(velDeltaDistance/walkingSpeed, 2), 2);
return percent;
//get the value on a declining sigmoid shaped curve that decays from 1 to 0
return Mathf.Pow(1 - Mathf.Pow(velDeltaMagnitude / walkingSpeed, 2), 2);
// public float VelocityInverseLerp(Vector3 velocityGoal)
// {
// avgVelValue = GetVelocity();
//
// velDeltaDistance = Vector3.Distance(avgVelValue, velocityGoal);
//// float percent = Mathf.InverseLerp(walkingSpeed, 0, velDeltaDistance);
// float percent = Mathf.InverseLerp(velocityGoal.magnitude, 0, velDeltaDistance);
// return percent;
// }
// public float VelocityInverseLerp(Vector3 velocityGoal, Vector3 currentVel)
// {
// avgVelValue = GetVelocity();
//
// velDeltaDistance = Vector3.Distance(currentVel, velocityGoal);
// float percent = Mathf.InverseLerp(walkingSpeed, 0, velDeltaDistance);
// return percent;
// }
// public float VelocityInverseLerp(Vector3 a, Vector3 b, Vector3 value)
// {
// Vector3 AB = b - a;
// Vector3 AV = value - a;
// return Vector3.Dot(AV, AB) / Vector3.Dot(AB, AB);
// }
// void FixedUpdate()
// {
// var cubeForward = orientationCube.transform.forward;
// orientationCube.UpdateOrientation(hips, target.transform);
// // Set reward for this step according to mixture of the following elements.
// // a. Velocity alignment with goal direction.
// var moveTowardsTargetReward = Vector3.Dot(cubeForward,
// Vector3.ClampMagnitude(m_JdController.bodyPartsDict[hips].rb.velocity, maximumWalkingSpeed));
// // b. Rotation alignment with goal direction.
// var lookAtTargetReward = Vector3.Dot(cubeForward, head.forward);
// // c. Encourage head height. //Should normalize to ~1
// var headHeightOverFeetReward =
// ((head.position.y - footL.position.y) + (head.position.y - footR.position.y) / 10);
// AddReward(
// + 0.02f * moveTowardsTargetReward
// + 0.02f * lookAtTargetReward
// + 0.005f * headHeightOverFeetReward
// );
// }
/// <summary>
/// Agent touched the target

{
SetTorsoMass();
}
}
}
撰写 预览
正在加载...
取消
保存