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
 
 
 
 
 

291 行
11 KiB

using UnityEngine;
using Unity.MLAgents;
using Unity.Barracuda;
using Unity.MLAgents.Actuators;
using Unity.MLAgentsExamples;
using Unity.MLAgents.Sensors;
using Random = UnityEngine.Random;
[RequireComponent(typeof(JointDriveController))] // Required to set joint forces
public class CrawlerAgent : Agent
{
[Header("Walk Speed")]
[Range(0.1f, m_maxWalkingSpeed)]
[SerializeField]
[Tooltip(
"The speed the agent will try to match.\n\n" +
"TRAINING:\n" +
"For VariableSpeed envs, this value will randomize at the start of each training episode.\n" +
"Otherwise the agent will try to match the speed set here.\n\n" +
"INFERENCE:\n" +
"During inference, VariableSpeed agents will modify their behavior based on this value " +
"whereas the CrawlerDynamic & CrawlerStatic agents will run at the speed specified during training "
)]
//The walking speed to try and achieve
private float m_TargetWalkingSpeed = m_maxWalkingSpeed;
const float m_maxWalkingSpeed = 15; //The max walking speed
//The current target walking speed. Clamped because a value of zero will cause NaNs
public float TargetWalkingSpeed
{
get { return m_TargetWalkingSpeed; }
set { m_TargetWalkingSpeed = Mathf.Clamp(value, .1f, m_maxWalkingSpeed); }
}
//The direction an agent will walk during training.
[Header("Target To Walk Towards")]
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")] [Space(10)] public Transform body;
public Transform leg0Upper;
public Transform leg0Lower;
public Transform leg1Upper;
public Transform leg1Lower;
public Transform leg2Upper;
public Transform leg2Lower;
public Transform leg3Upper;
public Transform leg3Lower;
//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;
[Header("Foot Grounded Visualization")]
[Space(10)]
public bool useFootGroundedVisualization;
public MeshRenderer foot0;
public MeshRenderer foot1;
public MeshRenderer foot2;
public MeshRenderer foot3;
public Material groundedMaterial;
public Material unGroundedMaterial;
public override void Initialize()
{
SpawnTarget(TargetPrefab, transform.position); //spawn target
m_OrientationCube = GetComponentInChildren<OrientationCubeController>();
m_DirectionIndicator = GetComponentInChildren<DirectionIndicator>();
m_JdController = GetComponent<JointDriveController>();
//Setup each body part
m_JdController.SetupBodyPart(body);
m_JdController.SetupBodyPart(leg0Upper);
m_JdController.SetupBodyPart(leg0Lower);
m_JdController.SetupBodyPart(leg1Upper);
m_JdController.SetupBodyPart(leg1Lower);
m_JdController.SetupBodyPart(leg2Upper);
m_JdController.SetupBodyPart(leg2Lower);
m_JdController.SetupBodyPart(leg3Upper);
m_JdController.SetupBodyPart(leg3Lower);
}
/// <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);
}
/// <summary>
/// Loop over body parts and reset them to initial conditions.
/// </summary>
public override void OnEpisodeBegin()
{
foreach (var bodyPart in m_JdController.bodyPartsDict.Values)
{
bodyPart.Reset(bodyPart);
}
//Random start rotation to help generalize
body.rotation = Quaternion.Euler(0, Random.Range(0.0f, 360.0f), 0);
UpdateOrientationObjects();
//Set our goal walking speed
TargetWalkingSpeed = Random.Range(0.1f, m_maxWalkingSpeed);
}
/// <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); // Is this bp touching the ground
if (bp.rb.transform != body)
{
sensor.AddObservation(bp.currentStrength / m_JdController.maxJointForceLimit);
}
}
/// <summary>
/// Loop over body parts to add them to observation.
/// </summary>
public override void CollectObservations(VectorSensor sensor)
{
var cubeForward = m_OrientationCube.transform.forward;
//velocity we want to match
var velGoal = cubeForward * TargetWalkingSpeed;
//ragdoll's avg vel
var avgVel = GetAvgVelocity();
//current ragdoll velocity. normalized
sensor.AddObservation(Vector3.Distance(velGoal, avgVel));
//avg body vel relative to cube
sensor.AddObservation(m_OrientationCube.transform.InverseTransformDirection(avgVel));
//vel goal relative to cube
sensor.AddObservation(m_OrientationCube.transform.InverseTransformDirection(velGoal));
//rotation delta
sensor.AddObservation(Quaternion.FromToRotation(body.forward, cubeForward));
//Add pos of target relative to orientation cube
sensor.AddObservation(m_OrientationCube.transform.InverseTransformPoint(m_Target.transform.position));
RaycastHit hit;
float maxRaycastDist = 10;
if (Physics.Raycast(body.position, Vector3.down, out hit, maxRaycastDist))
{
sensor.AddObservation(hit.distance / maxRaycastDist);
}
else
sensor.AddObservation(1);
foreach (var bodyPart in m_JdController.bodyPartsList)
{
CollectObservationBodyPart(bodyPart, sensor);
}
}
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 continuousActions = actionBuffers.ContinuousActions;
var i = -1;
// Pick a new target joint rotation
bpDict[leg0Upper].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[leg1Upper].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[leg2Upper].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[leg3Upper].SetJointTargetRotation(continuousActions[++i], continuousActions[++i], 0);
bpDict[leg0Lower].SetJointTargetRotation(continuousActions[++i], 0, 0);
bpDict[leg1Lower].SetJointTargetRotation(continuousActions[++i], 0, 0);
bpDict[leg2Lower].SetJointTargetRotation(continuousActions[++i], 0, 0);
bpDict[leg3Lower].SetJointTargetRotation(continuousActions[++i], 0, 0);
// Update joint strength
bpDict[leg0Upper].SetJointStrength(continuousActions[++i]);
bpDict[leg1Upper].SetJointStrength(continuousActions[++i]);
bpDict[leg2Upper].SetJointStrength(continuousActions[++i]);
bpDict[leg3Upper].SetJointStrength(continuousActions[++i]);
bpDict[leg0Lower].SetJointStrength(continuousActions[++i]);
bpDict[leg1Lower].SetJointStrength(continuousActions[++i]);
bpDict[leg2Lower].SetJointStrength(continuousActions[++i]);
bpDict[leg3Lower].SetJointStrength(continuousActions[++i]);
}
void FixedUpdate()
{
UpdateOrientationObjects();
// If enabled the feet will light up green when the foot is grounded.
// This is just a visualization and isn't necessary for function
if (useFootGroundedVisualization)
{
foot0.material = m_JdController.bodyPartsDict[leg0Lower].groundContact.touchingGround
? groundedMaterial
: unGroundedMaterial;
foot1.material = m_JdController.bodyPartsDict[leg1Lower].groundContact.touchingGround
? groundedMaterial
: unGroundedMaterial;
foot2.material = m_JdController.bodyPartsDict[leg2Lower].groundContact.touchingGround
? groundedMaterial
: unGroundedMaterial;
foot3.material = m_JdController.bodyPartsDict[leg3Lower].groundContact.touchingGround
? groundedMaterial
: unGroundedMaterial;
}
var cubeForward = m_OrientationCube.transform.forward;
// Set reward for this step according to mixture of the following elements.
// a. Match target speed
//This reward will approach 1 if it matches perfectly and approach zero as it deviates
var matchSpeedReward = GetMatchingVelocityReward(cubeForward * TargetWalkingSpeed, GetAvgVelocity());
// 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, body.forward) + 1) * .5F;
AddReward(matchSpeedReward * lookAtTargetReward);
}
/// <summary>
/// Update OrientationCube and DirectionIndicator
/// </summary>
void UpdateOrientationObjects()
{
m_OrientationCube.UpdateOrientation(body, m_Target);
if (m_DirectionIndicator)
{
m_DirectionIndicator.MatchOrientation(m_OrientationCube.transform);
}
}
/// <summary>
///Returns the average velocity of all of the body parts
///Using the velocity of the body only has shown to result in more erratic movement from the limbs
///Using the average helps prevent this erratic movement
/// </summary>
Vector3 GetAvgVelocity()
{
Vector3 velSum = Vector3.zero;
Vector3 avgVel = Vector3.zero;
//ALL RBS
int numOfRb = 0;
foreach (var item in m_JdController.bodyPartsList)
{
numOfRb++;
velSum += item.rb.velocity;
}
avgVel = velSum / numOfRb;
return avgVel;
}
/// <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, TargetWalkingSpeed);
//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 / TargetWalkingSpeed, 2), 2);
}
/// <summary>
/// Agent touched the target
/// </summary>
public void TouchedTarget()
{
AddReward(1f);
}
}