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
 
 
 
 
 

691 行
30 KiB

using System;
using System.Linq;
using System.Collections.Generic;
using System.Runtime.CompilerServices;
using UnityEngine;
using UnityEngine.Assertions;
using Unity.MLAgents.Sensors;
using UnityEngine.Profiling;
[assembly: InternalsVisibleTo("Unity.ML-Agents.Extensions.EditorTests")]
namespace Unity.MLAgents.Extensions.Sensors
{
/// <summary>
/// Enum describing what kind of depth type the data should be organized as
/// </summary>
public enum GridDepthType { Channel, ChannelHot, Counting };
/// <summary>
/// Grid-based sensor.
/// </summary>
public class GridSensor : ISensor, IBuiltInSensor
{
string m_Name;
Vector3 m_CellScale;
Vector3Int m_GridNum;
bool m_RotateWithAgent;
GameObject m_RootReference;
int m_MaxColliderBufferSize;
int m_InitialColliderBufferSize;
LayerMask m_ColliderMask;
GridDepthType m_GridDepthType;
int[] m_ChannelDepths;
string[] m_DetectableObjects;
SensorCompressionType m_CompressionType;
ObservationSpec m_ObservationSpec;
// Buffers
internal float[] m_PerceptionBuffer;
float[] m_ChannelHotDefaultPerceptionBuffer;
Color[] m_PerceptionColors;
Texture2D m_PerceptionTexture;
Collider[] m_ColliderBuffer;
float[] m_CellDataBuffer;
int[] m_ChannelOffsets;
Vector3[] m_CellLocalPositions;
int[] m_GizmoColorIndexes;
Vector3[] m_CellGlobalPosition;
// Utility Constants Calculated on Init
int m_NumCells;
int m_CellObservationSize;
float m_InverseSphereRadius;
float m_OffsetGridNumSide;
public GridSensor(
string name,
Vector3 cellScale,
Vector3Int gridNum,
bool rotateWithAgent,
int[] channelDepths,
string[] detectableObjects,
LayerMask colliderMask,
GridDepthType depthType,
GameObject rootReference,
SensorCompressionType compression,
int maxColliderBufferSize,
int initialColliderBufferSize
)
{
m_Name = name;
m_CellScale = cellScale;
m_GridNum = gridNum;
m_RotateWithAgent = rotateWithAgent;
m_RootReference = rootReference;
m_MaxColliderBufferSize = maxColliderBufferSize;
m_InitialColliderBufferSize = initialColliderBufferSize;
m_ColliderMask = colliderMask;
m_GridDepthType = depthType;
m_ChannelDepths = channelDepths;
m_DetectableObjects = detectableObjects;
m_CompressionType = compression;
if (m_GridNum.y != 1)
{
throw new UnityAgentsException("GridSensor only supports 2D grids.");
}
if (m_GridDepthType == GridDepthType.Counting && m_DetectableObjects.Length != m_ChannelDepths.Length)
{
throw new UnityAgentsException("The channels of a CountingGridSensor is equal to the number of detectableObjects");
}
InitGridParameters();
InitDepthType();
InitCellPoints();
ResetPerceptionBuffer();
m_ObservationSpec = ObservationSpec.Visual(m_GridNum.x, m_GridNum.z, m_CellObservationSize);
m_PerceptionTexture = new Texture2D(m_GridNum.x, m_GridNum.z, TextureFormat.RGB24, false);
m_ColliderBuffer = new Collider[Math.Min(m_MaxColliderBufferSize, m_InitialColliderBufferSize)];
}
public SensorCompressionType CompressionType
{
get { return m_CompressionType; }
set { m_CompressionType = value; }
}
public int[] GizmoColorIndexes
{
get { return m_GizmoColorIndexes; }
}
/// <summary>
/// Initializes the constant parameters used within the perceive method call
/// </summary>
void InitGridParameters()
{
m_NumCells = m_GridNum.x * m_GridNum.z;
float sphereRadiusX = (m_CellScale.x * m_GridNum.x) / Mathf.Sqrt(2);
float sphereRadiusZ = (m_CellScale.z * m_GridNum.z) / Mathf.Sqrt(2);
m_InverseSphereRadius = 1.0f / Mathf.Max(sphereRadiusX, sphereRadiusZ);
m_OffsetGridNumSide = (m_GridNum.z - 1f) / 2f;
}
/// <summary>
/// Initializes the constant parameters that are based on the Grid Depth Type
/// Sets the ObservationPerCell and the ChannelOffsets properties
/// </summary>
void InitDepthType()
{
if (m_GridDepthType == GridDepthType.ChannelHot)
{
m_CellObservationSize = m_ChannelDepths.Sum();
m_ChannelOffsets = new int[m_ChannelDepths.Length];
for (int i = 1; i < m_ChannelDepths.Length; i++)
{
m_ChannelOffsets[i] = m_ChannelOffsets[i - 1] + m_ChannelDepths[i - 1];
}
m_ChannelHotDefaultPerceptionBuffer = new float[m_CellObservationSize];
for (int i = 0; i < m_ChannelDepths.Length; i++)
{
if (m_ChannelDepths[i] > 1)
{
m_ChannelHotDefaultPerceptionBuffer[m_ChannelOffsets[i]] = 1;
}
}
}
else
{
m_CellObservationSize = m_ChannelDepths.Length;
}
// The maximum number of channels in the final output must be less than 255 * 3 because the "number of PNG images" to generate must fit in one byte
Assert.IsTrue(m_CellObservationSize < (255 * 3), "The maximum number of channels per cell must be less than 255 * 3");
}
/// <summary>
/// Initializes the location of the CellPoints property
/// </summary>
void InitCellPoints()
{
m_CellLocalPositions = new Vector3[m_NumCells];
for (int i = 0; i < m_NumCells; i++)
{
m_CellLocalPositions[i] = CellToLocalPoint(i);
}
}
/// <inheritdoc/>
public void Reset() { }
/// <summary>
/// Clears the perception buffer before loading in new data. If the gridDepthType is ChannelHot, then it initializes the
/// Reset() also reinits the cell activity array (for debug)
/// </summary>
public void ResetPerceptionBuffer()
{
if (m_PerceptionBuffer != null)
{
if (m_GridDepthType == GridDepthType.ChannelHot)
{
// Copy the default value to the array
for (int i = 0; i < m_NumCells; i++)
{
Array.Copy(m_ChannelHotDefaultPerceptionBuffer, 0, m_PerceptionBuffer, i * m_CellObservationSize, m_CellObservationSize);
}
}
else
{
Array.Clear(m_PerceptionBuffer, 0, m_PerceptionBuffer.Length);
}
}
else
{
m_PerceptionBuffer = new float[m_CellObservationSize * m_NumCells];
m_ColliderBuffer = new Collider[Math.Min(m_MaxColliderBufferSize, m_InitialColliderBufferSize)];
m_CellDataBuffer = new float[m_ChannelDepths.Length];
m_PerceptionColors = new Color[m_NumCells];
m_CellGlobalPosition = new Vector3[m_NumCells];
}
}
public void ResetGizmoBuffer()
{
// Ensure to init arrays if not yet assigned (for editor)
if (m_GizmoColorIndexes == null)
m_GizmoColorIndexes = new int[m_NumCells];
// Assign the default color to the cell activities
for (int i = 0; i < m_NumCells; i++)
{
m_GizmoColorIndexes[i] = -1;
}
}
/// <inheritdoc/>
public string GetName()
{
return m_Name;
}
/// <inheritdoc/>
public CompressionSpec GetCompressionSpec()
{
return new CompressionSpec(CompressionType);
}
/// <inheritdoc/>
public BuiltInSensorType GetBuiltInSensorType()
{
return BuiltInSensorType.GridSensor;
}
/// <summary>
/// GetCompressedObservation - Calls Perceive then puts the data stored on the perception buffer
/// onto the m_perceptionTexture2D to be converted to a byte array and returned
/// </summary>
/// <returns>byte[] containing the compressed observation of the grid observation</returns>
public byte[] GetCompressedObservation()
{
using (TimerStack.Instance.Scoped("GridSensor.GetCompressedObservation"))
{
var allBytes = new List<byte>();
var numImages = (m_CellObservationSize + 2) / 3;
for (int i = 0; i < numImages; i++)
{
var channelIndex = 3 * i;
ChannelsToTexture(channelIndex, Math.Min(3, m_CellObservationSize - channelIndex));
allBytes.AddRange(m_PerceptionTexture.EncodeToPNG());
}
return allBytes.ToArray();
}
}
/// <summary>
/// ChannelsToTexture - Takes the channel index and the numChannelsToAdd.
/// For each cell and for each channel to add, sets it to a value of the color specified for that cell.
/// All colors are then set to the perceptionTexture via SetPixels.
/// m_perceptionTexture2D can then be read as an image as it now contains all of the information that was
/// stored in the channels
/// </summary>
/// <param name="channelIndex"></param>
/// <param name="numChannelsToAdd"></param>
void ChannelsToTexture(int channelIndex, int numChannelsToAdd)
{
for (int i = 0; i < m_NumCells; i++)
{
for (int j = 0; j < numChannelsToAdd; j++)
{
m_PerceptionColors[i][j] = m_PerceptionBuffer[i * m_CellObservationSize + channelIndex + j];
}
}
m_PerceptionTexture.SetPixels(m_PerceptionColors);
}
/// <summary>
/// Perceive - Clears the buffers, calls overlap box on the actual cell (the actual perception part)
/// for all found colliders, LoadObjectData is called
/// </summary>
internal void Perceive()
{
if (m_ColliderBuffer == null)
{
return;
}
ResetPerceptionBuffer();
using (TimerStack.Instance.Scoped("GridSensor.Perceive"))
{
var halfCellScale = new Vector3(m_CellScale.x / 2f, m_CellScale.y, m_CellScale.z / 2f);
for (var cellIndex = 0; cellIndex < m_NumCells; cellIndex++)
{
var cellCenter = GetCellGlobalPosition(cellIndex);
var numFound = BufferResizingOverlapBoxNonAlloc(cellCenter, halfCellScale, GetGridRotation());
if (numFound > 0)
{
if (m_GridDepthType == GridDepthType.Counting)
{
ParseCollidersAll(m_ColliderBuffer, numFound, cellIndex, cellCenter);
}
else
{
ParseCollidersClosest(m_ColliderBuffer, numFound, cellIndex, cellCenter);
}
}
}
}
}
/// <summary>
/// This method attempts to perform the Physics.OverlapBoxNonAlloc and will double the size of the Collider buffer
/// if the number of Colliders in the buffer after the call is equal to the length of the buffer.
/// </summary>
/// <param name="cellCenter"></param>
/// <param name="halfCellScale"></param>
/// <param name="rotation"></param>
/// <returns></returns>
int BufferResizingOverlapBoxNonAlloc(Vector3 cellCenter, Vector3 halfCellScale, Quaternion rotation)
{
int numFound;
// Since we can only get a fixed number of results, requery
// until we're sure we can hold them all (or until we hit the max size).
while (true)
{
numFound = Physics.OverlapBoxNonAlloc(cellCenter, halfCellScale, m_ColliderBuffer, rotation, m_ColliderMask);
if (numFound == m_ColliderBuffer.Length && m_ColliderBuffer.Length < m_MaxColliderBufferSize)
{
m_ColliderBuffer = new Collider[Math.Min(m_MaxColliderBufferSize, m_ColliderBuffer.Length * 2)];
m_InitialColliderBufferSize = m_ColliderBuffer.Length;
}
else
{
break;
}
}
return numFound;
}
/// <summary>
/// Parses the array of colliders found within a cell. Finds the closest gameobject to the agent root reference within the cell
/// </summary>
/// <param name="foundColliders">Array of the colliders found within the cell</param>
/// <param name="numFound">Number of colliders found.</param>
/// <param name="cellIndex">The index of the cell</param>
/// <param name="cellCenter">The center position of the cell</param>
void ParseCollidersClosest(Collider[] foundColliders, int numFound, int cellIndex, Vector3 cellCenter)
{
Profiler.BeginSample("GridSensor.ParseColliders");
GameObject closestColliderGo = null;
var minDistanceSquared = float.MaxValue;
for (var i = 0; i < numFound; i++)
{
var currentColliderGo = foundColliders[i].gameObject;
// Continue if the current collider go is the root reference
if (ReferenceEquals(currentColliderGo, m_RootReference))
continue;
var closestColliderPoint = foundColliders[i].ClosestPointOnBounds(cellCenter);
var currentDistanceSquared = (closestColliderPoint - m_RootReference.transform.position).sqrMagnitude;
// Checks if our colliders contain a detectable object
var index = -1;
for (var ii = 0; ii < m_DetectableObjects.Length; ii++)
{
if (currentColliderGo.CompareTag(m_DetectableObjects[ii]))
{
index = ii;
break;
}
}
if (index > -1 && currentDistanceSquared < minDistanceSquared)
{
minDistanceSquared = currentDistanceSquared;
closestColliderGo = currentColliderGo;
}
}
if (!ReferenceEquals(closestColliderGo, null))
{
LoadObjectData(closestColliderGo, cellIndex, (float)Math.Sqrt(minDistanceSquared) * m_InverseSphereRadius);
}
Profiler.EndSample();
}
/// <summary>
/// For each collider, calls LoadObjectData on the gameobejct
/// </summary>
/// <param name="foundColliders">The array of colliders</param>
/// <param name="cellIndex">The cell index the collider is in</param>
/// <param name="cellCenter">the center of the cell the collider is in</param>
void ParseCollidersAll(Collider[] foundColliders, int numFound, int cellIndex, Vector3 cellCenter)
{
Profiler.BeginSample("GridSensor.ParseColliders");
GameObject currentColliderGo = null;
Vector3 closestColliderPoint = Vector3.zero;
for (int i = 0; i < numFound; i++)
{
currentColliderGo = foundColliders[i].gameObject;
// Continue if the current collider go is the root reference
if (currentColliderGo == m_RootReference)
continue;
closestColliderPoint = foundColliders[i].ClosestPointOnBounds(cellCenter);
LoadObjectData(currentColliderGo, cellIndex,
Vector3.Distance(closestColliderPoint, m_RootReference.transform.position) * m_InverseSphereRadius);
}
Profiler.EndSample();
}
/// <summary>
/// GetObjectData - returns an array of values that represent the game object
/// This is one of the few methods that one may need to override to get their required functionality
/// For instance, if one wants specific information about the current gameobject, they can use this method
/// to extract it and then return it in an array format.
/// </summary>
/// <returns>
/// A float[] containing the data that holds the representative information of the passed in gameObject
/// </returns>
/// <param name="currentColliderGo">The game object that was found colliding with a certain cell</param>
/// <param name="typeIndex">The index of the type (tag) of the gameObject.
/// (e.g., if this GameObject had the 3rd tag out of 4, type_index would be 2.0f)</param>
/// <param name="normalizedDistance">A float between 0 and 1 describing the ratio of
/// the distance currentColliderGo is compared to the edge of the gridsensor</param>
/// <example>
/// Here is an example of extenind GetObjectData to include information about a potential Rigidbody:
/// <code>
/// protected override float[] GetObjectData(GameObject currentColliderGo,
/// float type_index, float normalized_distance)
/// {
/// float[] channelValues = new float[ChannelDepth.Length]; // ChannelDepth.Length = 4 in this example
/// channelValues[0] = type_index;
/// Rigidbody goRb = currentColliderGo.GetComponent&lt;Rigidbody&gt;();
/// if (goRb != null)
/// {
/// channelValues[1] = goRb.velocity.x;
/// channelValues[2] = goRb.velocity.y;
/// channelValues[3] = goRb.velocity.z;
/// }
/// return channelValues;
/// }
/// </code>
/// </example>
protected virtual float[] GetObjectData(GameObject currentColliderGo, float typeIndex, float normalizedDistance)
{
Array.Clear(m_CellDataBuffer, 0, m_CellDataBuffer.Length);
m_CellDataBuffer[0] = typeIndex;
return m_CellDataBuffer;
}
/// <summary>
/// Runs basic validation assertions to check that the values can be normalized
/// </summary>
/// <param name="channelValues">The values to be validated</param>
/// <param name="currentColliderGo">The gameobject used for better error messages</param>
protected virtual void ValidateValues(float[] channelValues, GameObject currentColliderGo)
{
for (int j = 0; j < channelValues.Length; j++)
{
if (channelValues[j] < 0)
throw new UnityAgentsException("Expected ChannelValue[" + j + "] for " + currentColliderGo.name + " to be non-negative, was " + channelValues[j]);
if (channelValues[j] > m_ChannelDepths[j])
throw new UnityAgentsException("Expected ChannelValue[" + j + "] for " + currentColliderGo.name + " to be less than ChannelDepth[" + j + "] (" + m_ChannelDepths[j] + "), was " + channelValues[j]);
}
}
/// <summary>
/// LoadObjectData - If the GameObject matches a tag, GetObjectData is called to extract the data from the GameObject
/// then the data is transformed based on the GridDepthType of the gridsensor.
/// Further documetation on the GridDepthType can be found below
/// </summary>
/// <param name="currentColliderGo">The game object that was found colliding with a certain cell</param>
/// <param name="cellIndex">The index of the current cell</param>
/// <param name="normalizedDistance">A float between 0 and 1 describing the ratio of
/// the distance currentColliderGo is compared to the edge of the gridsensor</param>
protected virtual void LoadObjectData(GameObject currentColliderGo, int cellIndex, float normalizedDistance)
{
Profiler.BeginSample("GridSensor.LoadObjectData");
var channelHotVals = new ArraySegment<float>(m_PerceptionBuffer, cellIndex * m_CellObservationSize, m_CellObservationSize);
for (var i = 0; i < m_DetectableObjects.Length; i++)
{
if (m_GridDepthType != GridDepthType.Counting)
{
for (var ii = 0; ii < channelHotVals.Count; ii++)
{
m_PerceptionBuffer[channelHotVals.Offset + ii] = 0f;
}
}
if (!ReferenceEquals(currentColliderGo, null) && currentColliderGo.CompareTag(m_DetectableObjects[i]))
{
// TODO: Create the array already then set the values using "out" in GetObjectData
// Using i+1 as the type index as "0" represents "empty"
var channelValues = GetObjectData(currentColliderGo, (float)i + 1, normalizedDistance);
ValidateValues(channelValues, currentColliderGo);
switch (m_GridDepthType)
{
case GridDepthType.Channel:
{
// The observations are "channel based" so each grid is WxHxC where C is the number of channels
// This typically means that each channel value is normalized between 0 and 1
// If channelDepth is 1, the value is assumed normalized, else the value is normalized by the channelDepth
// The channels are then stored consecutively in PerceptionBuffer.
// NOTE: This is the only grid type that uses floating point values
// For example, if a cell contains the 3rd type of 5 possible on the 2nd team of 3 possible teams:
// channelValues = {2, 1}
// ObservationPerCell = channelValues.Length
// channelValues = {2f/5f, 1f/3f} = {.4, .33..}
// Array.Copy(channelValues, 0, PerceptionBuffer, cell_id*ObservationPerCell, ObservationPerCell);
for (int j = 0; j < channelValues.Length; j++)
{
channelValues[j] /= m_ChannelDepths[j];
}
Array.Copy(channelValues, 0, m_PerceptionBuffer, cellIndex * m_CellObservationSize, m_CellObservationSize);
break;
}
case GridDepthType.ChannelHot:
{
// The observations are "channel hot" so each grid is WxHxD where D is the sum of all of the channel depths
// The opposite of the "channel based" case, the channel values are represented as one hot vector per channel and then concatenated together
// Thus channelDepth is assumed to be greater than 1.
// For example, if a cell contains the 3rd type of 5 possible on the 2nd team of 3 possible teams,
// channelValues = {2, 1}
// channelOffsets = {5, 3}
// ObservationPerCell = 5 + 3 = 8
// channelHotVals = {0, 0, 1, 0, 0, 0, 1, 0}
// Array.Copy(channelHotVals, 0, PerceptionBuffer, cell_id*ObservationPerCell, ObservationPerCell);
for (int j = 0; j < channelValues.Length; j++)
{
if (m_ChannelDepths[j] > 1)
{
m_PerceptionBuffer[channelHotVals.Offset + (int)channelValues[j] + m_ChannelOffsets[j]] = 1f;
}
else
{
m_PerceptionBuffer[channelHotVals.Offset + m_ChannelOffsets[j]] = channelValues[j];
}
}
break;
}
case GridDepthType.Counting:
{
// The observations are "channel count" so each grid is WxHxC where C is the number of tags
// This means that each value channelValues[i] is a counter of gameobject included into grid cells
// where i is the index of the tag in DetectableObjects
int countIndex = cellIndex * m_CellObservationSize + i;
m_PerceptionBuffer[countIndex] = Mathf.Min(1f, m_PerceptionBuffer[countIndex] + 1f / m_ChannelDepths[i]);
break;
}
}
break;
}
}
Profiler.EndSample();
}
/// <summary>Converts the index of the cell to the 3D point (y is zero) relative to grid center</summary>
/// <returns>Vector3 of the position of the center of the cell relative to grid center</returns>
/// <param name="cell">The index of the cell</param>
Vector3 CellToLocalPoint(int cellIndex)
{
float x = (cellIndex % m_GridNum.z - m_OffsetGridNumSide) * m_CellScale.x;
float z = (cellIndex / m_GridNum.z - m_OffsetGridNumSide) * m_CellScale.z - (m_GridNum.z - m_GridNum.x);
return new Vector3(x, 0, z);
}
internal Vector3 GetCellGlobalPosition(int cellIndex)
{
if (m_RotateWithAgent)
{
return m_RootReference.transform.TransformPoint(m_CellLocalPositions[cellIndex]);
}
else
{
return m_CellLocalPositions[cellIndex] + m_RootReference.transform.position;
}
}
internal Quaternion GetGridRotation()
{
return m_RotateWithAgent ? m_RootReference.transform.rotation : Quaternion.identity;
}
/// <inheritdoc/>
public void Update()
{
using (TimerStack.Instance.Scoped("GridSensor.Update"))
{
Perceive();
}
}
/// <inheritdoc/>
public ObservationSpec GetObservationSpec()
{
return m_ObservationSpec;
}
/// <inheritdoc/>
public int Write(ObservationWriter writer)
{
using (TimerStack.Instance.Scoped("GridSensor.Write"))
{
int index = 0;
for (var h = m_GridNum.z - 1; h >= 0; h--)
{
for (var w = 0; w < m_GridNum.x; w++)
{
for (var d = 0; d < m_CellObservationSize; d++)
{
writer[h, w, d] = m_PerceptionBuffer[index];
index++;
}
}
}
return index;
}
}
internal int[] PerceiveGizmoColor()
{
ResetGizmoBuffer();
var halfCellScale = new Vector3(m_CellScale.x / 2f, m_CellScale.y, m_CellScale.z / 2f);
for (var cellIndex = 0; cellIndex < m_NumCells; cellIndex++)
{
var cellCenter = GetCellGlobalPosition(cellIndex);
var numFound = BufferResizingOverlapBoxNonAlloc(cellCenter, halfCellScale, GetGridRotation());
var minDistanceSquared = float.MaxValue;
var tagIndex = -1;
for (var i = 0; i < numFound; i++)
{
var currentColliderGo = m_ColliderBuffer[i].gameObject;
if (ReferenceEquals(currentColliderGo, m_RootReference))
continue;
var closestColliderPoint = m_ColliderBuffer[i].ClosestPointOnBounds(cellCenter);
var currentDistanceSquared = (closestColliderPoint - m_RootReference.transform.position).sqrMagnitude;
// Checks if our colliders contain a detectable object
var index = -1;
for (var ii = 0; ii < m_DetectableObjects.Length; ii++)
{
if (currentColliderGo.CompareTag(m_DetectableObjects[ii]))
{
index = ii;
break;
}
}
if (index > -1 && currentDistanceSquared < minDistanceSquared)
{
minDistanceSquared = currentDistanceSquared;
tagIndex = index;
}
}
m_GizmoColorIndexes[cellIndex] = tagIndex;
}
return m_GizmoColorIndexes;
}
internal Vector3[] GetGizmoPositions()
{
for (var i = 0; i < m_NumCells; i++)
{
m_CellGlobalPosition[i] = GetCellGlobalPosition(i);
}
return m_CellGlobalPosition;
}
}
}