Unity 机器学习代理工具包 (ML-Agents) 是一个开源项目,它使游戏和模拟能够作为训练智能代理的环境。
您最多选择25个主题 主题必须以中文或者字母或数字开头,可以包含连字符 (-),并且长度不得超过35个字符
 
 
 
 
 

211 行
7.9 KiB

using UnityEngine;
using Unity.MLAgents;
using Unity.MLAgents.Actuators;
using Unity.MLAgentsExamples;
using Unity.MLAgents.Sensors;
[RequireComponent(typeof(JointDriveController))] // Required to set joint forces
public class WormAgent : Agent
{
const float m_MaxWalkingSpeed = 10; //The max walking speed
[Header("Target Prefabs")] public Transform TargetPrefab; //Target prefab to use in Dynamic envs
private Transform m_Target; //Target the agent will walk towards during training.
[Header("Body Parts")] public Transform bodySegment0;
public Transform bodySegment1;
public Transform bodySegment2;
public Transform bodySegment3;
//This will be used as a stabilized model space reference point for observations
//Because ragdolls can move erratically during training, using a stabilized reference transform improves learning
OrientationCubeController m_OrientationCube;
//The indicator graphic gameobject that points towards the target
DirectionIndicator m_DirectionIndicator;
JointDriveController m_JdController;
private Vector3 m_StartingPos; //starting position of the agent
public override void Initialize()
{
SpawnTarget(TargetPrefab, transform.position); //spawn target
m_StartingPos = bodySegment0.position;
m_OrientationCube = GetComponentInChildren<OrientationCubeController>();
m_DirectionIndicator = GetComponentInChildren<DirectionIndicator>();
m_JdController = GetComponent<JointDriveController>();
UpdateOrientationObjects();
//Setup each body part
m_JdController.SetupBodyPart(bodySegment0);
m_JdController.SetupBodyPart(bodySegment1);
m_JdController.SetupBodyPart(bodySegment2);
m_JdController.SetupBodyPart(bodySegment3);
}
/// <summary>
/// Spawns a target prefab at pos
/// </summary>
/// <param name="prefab"></param>
/// <param name="pos"></param>
void SpawnTarget(Transform prefab, Vector3 pos)
{
m_Target = Instantiate(prefab, pos, Quaternion.identity, transform.parent);
}
/// <summary>
/// Loop over body parts and reset them to initial conditions.
/// </summary>
public override void OnEpisodeBegin()
{
foreach (var bodyPart in m_JdController.bodyPartsList)
{
bodyPart.Reset(bodyPart);
}
//Random start rotation to help generalize
bodySegment0.rotation = Quaternion.Euler(0, Random.Range(0.0f, 360.0f), 0);
UpdateOrientationObjects();
}
/// <summary>
/// Add relevant information on each body part to observations.
/// </summary>
public void CollectObservationBodyPart(BodyPart bp, VectorSensor sensor)
{
//GROUND CHECK
sensor.AddObservation(bp.groundContact.touchingGround ? 1 : 0); // Whether the bp touching the ground
//Get velocities in the context of our orientation cube's space
//Note: You can get these velocities in world space as well but it may not train as well.
sensor.AddObservation(m_OrientationCube.transform.InverseTransformDirection(bp.rb.velocity));
sensor.AddObservation(m_OrientationCube.transform.InverseTransformDirection(bp.rb.angularVelocity));
if (bp.rb.transform != bodySegment0)
{
//Get position relative to hips in the context of our orientation cube's space
sensor.AddObservation(
m_OrientationCube.transform.InverseTransformDirection(bp.rb.position - bodySegment0.position));
sensor.AddObservation(bp.rb.transform.localRotation);
}
if (bp.joint)
sensor.AddObservation(bp.currentStrength / m_JdController.maxJointForceLimit);
}
public override void CollectObservations(VectorSensor sensor)
{
RaycastHit hit;
float maxDist = 10;
if (Physics.Raycast(bodySegment0.position, Vector3.down, out hit, maxDist))
{
sensor.AddObservation(hit.distance / maxDist);
}
else
sensor.AddObservation(1);
var cubeForward = m_OrientationCube.transform.forward;
var velGoal = cubeForward * m_MaxWalkingSpeed;
sensor.AddObservation(m_OrientationCube.transform.InverseTransformDirection(velGoal));
sensor.AddObservation(Quaternion.Angle(m_OrientationCube.transform.rotation,
m_JdController.bodyPartsDict[bodySegment0].rb.rotation) / 180);
sensor.AddObservation(Quaternion.FromToRotation(bodySegment0.forward, cubeForward));
//Add pos of target relative to orientation cube
sensor.AddObservation(m_OrientationCube.transform.InverseTransformPoint(m_Target.transform.position));
foreach (var bodyPart in m_JdController.bodyPartsList)
{
CollectObservationBodyPart(bodyPart, sensor);
}
}
/// <summary>
/// Agent touched the target
/// </summary>
public void TouchedTarget()
{
AddReward(1f);
}
public override void OnActionReceived(ActionBuffers actionBuffers)
{
// The dictionary with all the body parts in it are in the jdController
var bpDict = m_JdController.bodyPartsDict;
var i = -1;
var continuousActions = actionBuffers.ContinuousActions;
// Pick a new target joint rotation
bpDict[bodySegment0].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[bodySegment1].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[bodySegment2].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
// Update joint strength
bpDict[bodySegment0].SetJointStrength(continuousActions[++i]);
bpDict[bodySegment1].SetJointStrength(continuousActions[++i]);
bpDict[bodySegment2].SetJointStrength(continuousActions[++i]);
//Reset if Worm fell through floor;
if (bodySegment0.position.y < m_StartingPos.y - 2)
{
EndEpisode();
}
}
void FixedUpdate()
{
UpdateOrientationObjects();
var velReward =
GetMatchingVelocityReward(m_OrientationCube.transform.forward * m_MaxWalkingSpeed,
m_JdController.bodyPartsDict[bodySegment0].rb.velocity);
//Angle of the rotation delta between cube and body.
//This will range from (0, 180)
var rotAngle = Quaternion.Angle(m_OrientationCube.transform.rotation,
m_JdController.bodyPartsDict[bodySegment0].rb.rotation);
//The reward for facing the target
var facingRew = 0f;
//If we are within 30 degrees of facing the target
if (rotAngle < 30)
{
//Set normalized facingReward
//Facing the target perfectly yields a reward of 1
facingRew = 1 - (rotAngle / 180);
}
//Add the product of these two rewards
AddReward(velReward * facingRew);
}
/// <summary>
/// Normalized value of the difference in actual speed vs goal walking speed.
/// </summary>
public float GetMatchingVelocityReward(Vector3 velocityGoal, Vector3 actualVelocity)
{
//distance between our actual velocity and goal velocity
var velDeltaMagnitude = Mathf.Clamp(Vector3.Distance(actualVelocity, velocityGoal), 0, m_MaxWalkingSpeed);
//return the value on a declining sigmoid shaped curve that decays from 1 to 0
//This reward will approach 1 if it matches perfectly and approach zero as it deviates
return Mathf.Pow(1 - Mathf.Pow(velDeltaMagnitude / m_MaxWalkingSpeed, 2), 2);
}
/// <summary>
/// Update OrientationCube and DirectionIndicator
/// </summary>
void UpdateOrientationObjects()
{
m_OrientationCube.UpdateOrientation(bodySegment0, m_Target);
if (m_DirectionIndicator)
{
m_DirectionIndicator.MatchOrientation(m_OrientationCube.transform);
}
}
}