def main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# system
gpuId = 0
# objects
objectHeight = [0.007, 0.013]
objectRadius = [0.030, 0.045]
nObjects = 1000
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.60
viewWorkspace = [(-1.0, 1.0), (-1.0, 1.0), (-1.0, 1.0)]
# visualization/saving
showViewer = False
showSteps = False
plotImages = False
# INITIALIZATION =================================================================================
rlEnv = RlEnvironment(showViewer, removeTable=True)
rlAgent = RlAgent(rlEnv, gpuId)
# RUN TEST =======================================================================================
for objIdx in xrange(nObjects):
obj = rlEnv.PlaceCylinderAtOrigin(objectHeight, objectRadius,
"cylinder-{}".format(objIdx), True)
cloud, normals = rlAgent.GetFullCloudAndNormals(
viewCenter, viewKeepout, viewWorkspace)
point_cloud.SaveMat("cylinder-{}.mat".format(objIdx), cloud, normals)
rlAgent.PlotCloud(cloud)
if plotImages:
point_cloud.Plot(cloud, normals, 2)
if showSteps:
raw_input("Placed cylinder-{}.".format(objIdx))
rlEnv.RemoveObjectSet([obj])
def main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# system
gpuId = 0
# objects
objectScale = [0.09, 0.17]
nObjects = 1000
directory = "/home/mgualti/Data/3DNet/Cat200_ModelDatabase/plate/"
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.60
viewWorkspace = [(-1.0, 1.0), (-1.0, 1.0), (-1.0, 1.0)]
# visualization/saving
showViewer = False
showSteps = False
plotImages = False
# INITIALIZATION =================================================================================
rlEnv = RlEnvironment(showViewer, removeTable=True)
rlAgent = RlAgent(rlEnv, gpuId)
# RUN TEST =======================================================================================
for objIdx in xrange(nObjects):
obj = rlEnv.Place3DNetObjectAtOrigin(directory, objectScale,
"plate-{}".format(objIdx), True)
cloud, normals = rlAgent.GetFullCloudAndNormals(
viewCenter, viewKeepout, viewWorkspace, False)
point_cloud.SaveMat("plate-{}.mat".format(objIdx), cloud, normals)
rlAgent.PlotCloud(cloud)
if plotImages:
point_cloud.Plot(cloud, normals, 2)
if showSteps:
raw_input("Placed plate-{}.".format(objIdx))
rlEnv.RemoveObjectSet([obj])
def __init__(self, rlEnvironment, gpuId, hasHistory, nSamples):
'''Initializes agent in the given environment.'''
RlAgent.__init__(self, rlEnvironment, gpuId)
# parameters
self.nSamples = nSamples
self.hasHistory = hasHistory
self.emptyStateImage = zeros((3, 60, 60), dtype='float')
self.emptyStateVector = zeros(6, dtype='float')
self.emptyState = [self.emptyStateImage, self.emptyStateVector]
self.minValue = -float('inf'); self.maxValue = float('inf')
# network parameters
self.conv1aOutputs = 12 # LeNet 20 outputs
self.conv1sOutputs = 12 # LeNet 20 outputs
self.conv2aOutputs = 24 # LeNet 50 outputs
self.conv2sOutputs = 24 # LeNet 50 outputs
self.ip1aOutputs = 200 # LeNet 500 outputs
self.ip1sOutputs = 200 # LeNet 500 outputs
self.ip1Outputs = 50 # MarcNet 60 outputs
self.ip2Outputs = 50 # MarcNet 60 outputs
def main():
'''Entrypoint to the program.
- Input objectClass: Folder in 3D Net database.
'''
# PARAMETERS =====================================================================================
# INITIALIZATION =================================================================================
rlEnv = RlEnvironment(True)
rlAgent = RlAgent(rlEnv)
# RUN TEST =======================================================================================
for placeIdx, placePose in enumerate(rlAgent.placePoses):
print("Showing pose {}: ".format(placeIdx))
print placePose
rlAgent.MoveHandToPose(placePose)
raw_input("Press [Enter] to continue...")
print("Finished.")
def Main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
params = loadmat("parameters.mat", squeeze_me=True)
randomSeed = params["randomSeed"]
tMax = params["tMax"]
nEpisodes = params["nEpisodes"]
trainEvery = params["trainEvery"]
unbiasOnEpisode = params["unbiasOnEpisode"]
saveFileName = params["saveFileName"]
loadNetwork = params["loadNetwork"]
loadDatabase = params["loadDatabase"]
showSteps = params["showSteps"]
# INITIALIZATION =================================================================================
# set random seeds
seed(randomSeed)
tensorflow.random.set_seed(randomSeed)
# initialize agent and environment
rlEnv = RlEnvironmentBottlesOnCoasters(params)
rlAgent = RlAgent(params)
# if testing, load previous results
if loadNetwork:
rlAgent.LoadQFunction()
if loadDatabase:
rlAgent.LoadExperienceDatabase()
# RUN TEST =======================================================================================
episodeReturn = []
nPlacedObjects = []
nGraspedObjects = []
episodeTime = []
timeStepEpsilon = []
databaseSize = []
losses = []
for episode in xrange(nEpisodes):
startTime = time()
# place random object in random orientation on table
rlEnv.MoveHandToHoldingPose()
rlEnv.PlaceObjects(False)
rlEnv.PlaceObjects(True)
if showSteps: raw_input("Placed objects.")
R = 0
nPlaced = 0
nGrasped = 0
holdingDesc = None
observations = []
actions = []
rewards = []
isGrasp = []
for t in xrange(tMax):
# get a point cloud
cloud = rlEnv.GetArtificialCloud()
#rlEnv.PlotCloud(cloud)
#if showSteps: raw_input("Acquired cloud.")
isGrasp.append(holdingDesc is None)
# get the next action
o, a, overtDesc, epsilon = rlAgent.SenseAndAct(
holdingDesc, cloud, t, rlEnv, episode >= unbiasOnEpisode)
# perform transition
holdingDesc, r = rlEnv.Transition(overtDesc, cloud)
# save experiences
timeStepEpsilon.append(epsilon)
observations.append(o)
actions.append(a)
rewards.append(r)
R += r
# compute task success -- number of objects placed
if isGrasp[-1]:
if holdingDesc is not None:
nGrasped += 1
if r < 0:
nPlaced -= 1
else:
if r == 1:
nPlaced += 1
rlAgent.AddExperienceMonteCarlo(observations, actions, rewards,
isGrasp)
# cleanup episode
rlEnv.ResetEpisode()
print("Episode {} had return {}".format(episode, R))
# training
if episode % trainEvery == trainEvery - 1:
losses.append(rlAgent.UpdateQFunctionMonteCarlo())
rlAgent.SaveQFunction()
# save results
episodeReturn.append(R)
nPlacedObjects.append(nPlaced)
nGraspedObjects.append(nGrasped)
episodeTime.append(time() - startTime)
databaseSize.append(rlAgent.GetNumberOfExperiences())
if episode % trainEvery == trainEvery - 1 or episode == nEpisodes - 1:
saveData = {
"episodeReturn": episodeReturn,
"nPlacedObjects": nPlacedObjects,
"nGraspedObjects": nGraspedObjects,
"episodeTime": episodeTime,
"timeStepEpsilon": timeStepEpsilon,
"databaseSize": databaseSize,
"losses": losses
}
saveData.update(params)
savemat(saveFileName, saveData)
# backup agent data
if episode == nEpisodes - 1:
rlAgent.SaveExperienceDatabase()
def main(saveFileSuffix):
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# objects
objectClass = "mug_train"
randomScale = True
targetObjectAxis = array([0, 0, 1])
maxAngleFromObjectAxis = 20 * (pi / 180)
maxObjectTableGap = 0.03
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.50
viewWorkspace = [(-1, 1), (-1, 1), (0.002, 1)]
# grasps
graspDetectMode = 0 # 0=sample, 1=sample+label
nGraspSamples = 100
graspScoreThresh = 350
# learning
nValueIterations = 70
nDataIterations = 50
nGraspIterations = 20
pickEpsilon = 1.0
placeEpsilon = 1.0
minPickEpsilon = 0.10
minPlaceEpsilon = 0.10
pickEpsilonDelta = 0.05
placeEpsilonDelta = 0.05
maxExperiences = 25000
trainingBatchSize = 25000
unbiasOnIteration = nValueIterations - 5
# visualization/saving
saveFileName = "results" + saveFileSuffix + ".mat"
recordLoss = True
showViewer = False
showSteps = False
# INITIALIZATION =================================================================================
threeDNet = ThreeDNet()
rlEnv = RlEnvironment(showViewer)
rlAgent = RlAgent(rlEnv)
nPlaceOptions = len(rlAgent.placePoses)
experienceDatabase = []
# RUN TEST =======================================================================================
averageReward = []
placeActionCounts = []
trainLosses = []
testLosses = []
databaseSize = []
iterationTime = []
for valueIterationIdx in xrange(nValueIterations):
print("Iteration {}. Epsilon pick: {}, place: {}".format(\
valueIterationIdx, pickEpsilon, placeEpsilon))
# 1. Collect data for this training iteration.
iterationStartTime = time.time()
R = []
placeCounts = zeros(nPlaceOptions)
# check if it's time to unbias data
if valueIterationIdx >= unbiasOnIteration:
maxExperiences = trainingBatchSize # selects all recent experiences, unbiased
pickEpsilon = 0 # estimating value function of actual policy
placeEpsilon = 0 # estimating value function of actual policy
for dataIterationIdx in xrange(nDataIterations):
# place random object in random orientation on table
fullObjName, objScale = threeDNet.GetRandomObjectFromClass(
objectClass, randomScale)
objHandle, objRandPose = rlEnv.PlaceObjectRandomOrientation(
fullObjName, objScale)
# move the hand to view position and capture a point cloud
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.PlotCloud(cloud)
# detect grasps in the sensory data
grasps = rlAgent.DetectGrasps(cloud, viewPoints, viewPointIndices,
nGraspSamples, graspScoreThresh,
graspDetectMode)
rlAgent.PlotGrasps(grasps)
if showSteps:
raw_input("Acquired grasps.")
if len(grasps) == 0:
print("No grasps found. Skipping iteration.")
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotCloud()
continue
for graspIterationIdx in xrange(nGraspIterations):
print("Episode {}.{}.{}.".format(valueIterationIdx,
dataIterationIdx,
graspIterationIdx))
# perform pick action
grasp = rlAgent.GetGrasp(grasps, pickEpsilon)
s = rlEnv.GetState(rlAgent, grasp, None)
rlAgent.PlotGrasps([grasp])
if showSteps:
print("Selected grasp.")
# perform place action
P = rlAgent.GetPlacePose(grasp, placeEpsilon)
rlAgent.MoveHandToPose(P)
ss = rlEnv.GetState(rlAgent, grasp, P)
rlAgent.MoveObjectToHandAtGrasp(grasp, objHandle)
r = rlEnv.RewardBinary(objHandle, targetObjectAxis,
maxAngleFromObjectAxis,
maxObjectTableGap)
print("The robot receives {} reward.".format(r))
if showSteps:
raw_input("Press [Enter] to continue...")
# add experience to database
experienceDatabase.append((s, ss, 0)) # grasp -> placement
experienceDatabase.append((ss, None, r)) # placement -> end
# record save data
R.append(r)
placeCounts += ss[1][len(s[1]) - nPlaceOptions:]
# cleanup this grasp iteration
rlAgent.UnplotGrasps()
rlEnv.MoveObjectToPose(objHandle, objRandPose)
# cleanup this data iteration
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotCloud()
# 2. Compute value labels for data.
experienceDatabase = rlAgent.PruneDatabase(experienceDatabase,
maxExperiences)
Dl = rlAgent.DownsampleAndLabelData(\
experienceDatabase, trainingBatchSize)
databaseSize.append(len(experienceDatabase))
# 3. Train network from replay database.
trainLoss, testLoss = rlAgent.Train(Dl, recordLoss=recordLoss)
trainLosses.append(trainLoss)
testLosses.append(testLoss)
pickEpsilon -= pickEpsilonDelta
placeEpsilon -= placeEpsilonDelta
pickEpsilon = max(minPickEpsilon, pickEpsilon)
placeEpsilon = max(minPlaceEpsilon, placeEpsilon)
# 4. Save results
averageReward.append(mean(R))
placeActionCounts.append(placeCounts)
iterationTime.append(time.time() - iterationStartTime)
saveData = {
"objectClass": objectClass,
"nGraspSamples": nGraspSamples,
"graspScoreThresh": graspScoreThresh,
"nValueIterations": nValueIterations,
"nDataIterations": nDataIterations,
"nGraspIterations": nGraspIterations,
"pickEpsilon": pickEpsilon,
"placeEpsilon": placeEpsilon,
"minPickEpsilon": minPickEpsilon,
"minPlaceEpsilon": minPlaceEpsilon,
"pickEpsilonDelta": pickEpsilonDelta,
"placeEpsilonDelta": placeEpsilonDelta,
"maxExperiences": maxExperiences,
"trainingBatchSize": trainingBatchSize,
"averageReward": averageReward,
"placeActionCounts": placeActionCounts,
"trainLoss": trainLosses,
"testLoss": testLosses,
"databaseSize": databaseSize,
"iterationTime": iterationTime,
"placePoses": rlAgent.placePoses
}
savemat(saveFileName, saveData)
def main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# objects
objectClass = "mug_train"
randomScale = True
targetObjectAxis = array([0, 0, 1])
maxAngleFromObjectAxis = 20 * (pi / 180)
maxObjectTableGap = 0.03
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.50
viewWorkspace = [(-1, 1), (-1, 1), (-1, 1)]
# grasps
graspDetectMode = 1 # 0=sample, 1=sample+label
nGraspSamples = 200
graspScoreThresh = 300
# learning
nTrainingIterations = 100
nEpisodes = 100
nReuses = 10
maxTimesteps = 10
gamma = 0.98
epsilon = 1.0
epsilonDelta = 0.05
minEpsilon = 0.05
maxExperiences = 50000
trainingBatchSize = 50000
unbiasOnIteration = nTrainingIterations - 5
# visualization/saving
saveFileName = "results.mat"
recordLoss = True
showViewer = False
showSteps = False
# INITIALIZATION =================================================================================
threeDNet = ThreeDNet()
rlEnv = RlEnvironment(showViewer)
rlAgent = RlAgent(rlEnv)
nPlaceOptions = len(rlAgent.placePoses)
experienceDatabase = []
# RUN TEST =======================================================================================
avgReturn = []
avgGraspsDetected = []
avgTopGraspsDetected = []
placeHistograms = []
avgGoodTempPlaceCount = []
avgBadTempPlaceCount = []
avgGoodFinalPlaceCount = []
avgBadFinalPlaceCount = []
trainLosses = []
testLosses = []
databaseSize = []
iterationTime = []
for trainingIteration in xrange(nTrainingIterations):
# initialization
iterationStartTime = time.time()
print("Iteration: {}, Epsilon: {}".format(trainingIteration, epsilon))
placeHistogram = zeros(nPlaceOptions)
Return = []
graspsDetected = []
topGraspsDetected = []
goodTempPlaceCount = []
badTempPlaceCount = []
goodFinalPlaceCount = []
badFinalPlaceCount = []
# check if it's time to unbias data
if trainingIteration >= unbiasOnIteration:
maxExperiences = trainingBatchSize # selects all recent experiences, unbiased
epsilon = 0 # estimating value function of actual policy
# for each episode/object placement
for episode in xrange(nEpisodes):
# place random object in random orientation on table
fullObjName, objScale = threeDNet.GetRandomObjectFromClass(
objectClass, randomScale)
objHandle, objRandPose = rlEnv.PlaceObjectRandomOrientation(
fullObjName, objScale)
# move the hand to view position(s) and capture a point cloud
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
# detect grasps in the sensor data
grasps = rlAgent.DetectGrasps(cloud, viewPoints, viewPointIndices,
nGraspSamples, graspScoreThresh,
graspDetectMode)
graspsStart = grasps
graspsDetected.append(len(grasps))
topGraspsCount = CountObjectTopGrasps(grasps, objRandPose,
maxAngleFromObjectAxis)
if len(grasps) == 0:
print("No grasps found. Skipping iteration.")
rlEnv.RemoveObject(objHandle)
continue
rlAgent.PlotCloud(cloud)
rlAgent.PlotGrasps(grasps)
for reuse in xrange(nReuses):
print("Episode {}.{}.{}.".format(trainingIteration, episode,
reuse))
if showSteps:
raw_input(
"Beginning of episode. Press [Enter] to continue...")
# initialize recording variables
episodePlaceHistogram = zeros(nPlaceOptions)
episodeReturn = 0
episodeGoodTempPlaceCount = 0
episodeBadTempPlaceCount = 0
episodeGoodFinalPlaceCount = 0
episodeBadFinalPlaceCount = 0
graspDetectionFailure = False
episodeExperiences = []
# initial state and first action
s, selectedGrasp = rlEnv.GetInitialState(rlAgent)
a, grasp, place = rlAgent.ChooseAction(s, grasps, epsilon)
rlAgent.PlotGrasps([grasp])
# for each time step in the episode
for t in xrange(maxTimesteps):
ss, selectedGrasp, rr = rlEnv.Transition(
rlAgent, objHandle, s, selectedGrasp, a, grasp, place,
targetObjectAxis, maxAngleFromObjectAxis,
maxObjectTableGap)
ssIsPlacedTempGood = ss[1][1]
ssIsPlacedTempBad = ss[1][2]
ssIsPlacedFinalGood = ss[1][3]
ssIsPlacedFinalBad = ss[1][4]
if showSteps:
raw_input(
"Transition {}. Press [Enter] to continue...".
format(t))
# re-detect only if a non-terminal placement just happened
if ssIsPlacedTempGood and place is not None:
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
grasps = rlAgent.DetectGrasps(cloud, viewPoints,
viewPointIndices,
nGraspSamples,
graspScoreThresh,
graspDetectMode)
graspsDetected.append(len(grasps))
topGraspsCount = CountObjectTopGrasps(
grasps, rlEnv.GetObjectPose(objHandle),
maxAngleFromObjectAxis)
topGraspsDetected.append(topGraspsCount)
if len(grasps) == 0:
print("Grasp detection failure.")
graspDetectionFailure = True
break
rlAgent.PlotCloud(cloud)
rlAgent.PlotGrasps(grasps)
# get next action
aa, ggrasp, pplace = rlAgent.ChooseAction(
ss, grasps, epsilon)
if ggrasp is not None: rlAgent.PlotGrasps([ggrasp])
if showSteps:
raw_input(
"Action {}. Press [Enter] to continue...".format(
t))
# add to database and record data
episodeExperiences.append((s, a, rr, ss, aa))
episodeReturn += (gamma**t) * rr
if place is not None:
episodeGoodTempPlaceCount += ssIsPlacedTempGood
episodeBadTempPlaceCount += ssIsPlacedTempBad
episodeGoodFinalPlaceCount += ssIsPlacedFinalGood
episodeBadFinalPlaceCount += ssIsPlacedFinalBad
placeHistogram += a[1]
# prepare for next time step
if ssIsPlacedTempBad or ssIsPlacedFinalGood or ssIsPlacedFinalBad:
break
s = ss
a = aa
grasp = ggrasp
place = pplace
# cleanup this reuse
if not graspDetectionFailure:
experienceDatabase += episodeExperiences
placeHistogram += episodePlaceHistogram
Return.append(episodeReturn)
goodTempPlaceCount.append(episodeGoodTempPlaceCount)
badTempPlaceCount.append(episodeBadTempPlaceCount)
goodFinalPlaceCount.append(episodeGoodFinalPlaceCount)
badFinalPlaceCount.append(episodeBadFinalPlaceCount)
rlEnv.MoveObjectToPose(objHandle, objRandPose)
grasps = graspsStart
# cleanup this episode
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotGrasps()
rlAgent.UnplotCloud()
# 2. Compute value labels for data.
experienceDatabase = rlAgent.PruneDatabase(experienceDatabase,
maxExperiences)
Dl = rlAgent.DownsampleAndLabelData(experienceDatabase,
trainingBatchSize, gamma)
databaseSize.append(len(experienceDatabase))
# 3. Train network from replay database.
trainLoss, testLoss = rlAgent.Train(Dl, recordLoss=recordLoss)
trainLosses.append(trainLoss)
testLosses.append(testLoss)
epsilon -= epsilonDelta
epsilon = max(minEpsilon, epsilon)
# 4. Save results
avgReturn.append(mean(Return))
avgGraspsDetected.append(mean(graspsDetected))
avgTopGraspsDetected.append(mean(topGraspsDetected))
placeHistograms.append(placeHistogram)
avgGoodTempPlaceCount.append(mean(goodTempPlaceCount))
avgBadTempPlaceCount.append(mean(badTempPlaceCount))
avgGoodFinalPlaceCount.append(mean(goodFinalPlaceCount))
avgBadFinalPlaceCount.append(mean(badFinalPlaceCount))
iterationTime.append(time.time() - iterationStartTime)
saveData = {
"objectClass": objectClass,
"randomScale": randomScale,
"maxAngleFromObjectAxis": maxAngleFromObjectAxis,
"maxObjectTableGap": maxObjectTableGap,
"nGraspSamples": nGraspSamples,
"graspScoreThresh": graspScoreThresh,
"graspDetectMode": graspDetectMode,
"nTrainingIterations": nTrainingIterations,
"nEpisodes": nEpisodes,
"maxTimesteps": maxTimesteps,
"gamma": gamma,
"epsilon": epsilon,
"minEpsilon": minEpsilon,
"epsilonDelta": epsilonDelta,
"maxExperiences": maxExperiences,
"trainingBatchSize": trainingBatchSize,
"avgReturn": avgReturn,
"avgGraspsDetected": avgGraspsDetected,
"avgTopGraspsDetected": avgTopGraspsDetected,
"placeHistograms": placeHistograms,
"avgGoodTempPlaceCount": avgGoodTempPlaceCount,
"avgBadTempPlaceCount": avgBadTempPlaceCount,
"avgGoodFinalPlaceCount": avgGoodFinalPlaceCount,
"avgBadFinalPlaceCount": avgBadFinalPlaceCount,
"trainLoss": trainLosses,
"testLoss": testLosses,
"databaseSize": databaseSize,
"iterationTime": iterationTime,
"placePoses": rlAgent.placePoses
}
savemat(saveFileName, saveData)
time.sleep(2)
print("Before 1,000 iterations have passed,\nyou will notice that the agent is getting better\nas it needs fewer steps to complete a corse.")
time.sleep(5)
print("When it has reached 5,000 iterations, you\nget to see exactly how it acts.")
time.sleep(5)
if __name__ == '__main__':
size = 10
epsilon = .5
bufferSize = 1000000
syncRate = 1000
batchSize = 32
discount = .98
graphics = False
environment = Environment(size,size)
agent = RlAgent(size,size,epsilon,bufferSize,syncRate,batchSize,discount)
i = 0
while True:
if i ==200:
agent.epsilon = .1
elif i == 1000:
agent.epsilon = .05
elif i == 5000:
agent.epsilon = .02
graphics = True
steps = 0
environment = Environment(size,size)
while environment.state is not None:
steps+=1
state1 = environment.state.copy()
def main(objectClass, epsilon):
'''Entrypoint to the program.
- Input objectClass: Folder in 3D Net database.
'''
# PARAMETERS =====================================================================================
# objects
randomScale = True
targetObjectAxis = array([0,0,1])
maxAngleFromObjectAxis = 20*(pi/180)
maxObjectTableGap = 0.02
# view
viewCenter = array([0,0,0])
viewKeepout = 0.50
viewWorkspace = [(-1,1),(-1,1),(-1,1)]
# grasps
graspDetectMode = 1 # 0=sample, 1=sample+label
nGraspSamples = 200
graspScoreThresh = 300
nGraspInliers = 3
# learning
weightsFileName = "/home/mgualti/mgualti/PickAndPlace/simulation/caffe/image_iter_5000.caffemodel"
nDataIterations = 300
# visualization/saving
showViewer = False
showEveryStep = False
saveFileName = "results-single-" + objectClass + "-epsilon" + str(epsilon) + ".mat"
# INITIALIZATION =================================================================================
threeDNet = ThreeDNet()
rlEnv = RlEnvironment(showViewer)
rlAgent = RlAgent(rlEnv)
if epsilon < 1.0: rlAgent.LoadNetworkWeights(weightsFileName)
Return = []
# RUN TEST =======================================================================================
# Collect data for this training iteration.
for dataIterationIdx in xrange(nDataIterations):
print("Iteration {}.".format(dataIterationIdx))
# place object in random orientation on table
fullObjName, objScale = threeDNet.GetRandomObjectFromClass(objectClass, randomScale)
objHandle, objRandPose = rlEnv.PlaceObjectRandomOrientation(fullObjName, objScale)
# move the hand to view position and capture a point cloud
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.PlotCloud(cloud)
# detect grasps in the sensory data
grasps = rlAgent.DetectGrasps(cloud, viewPoints, viewPointIndices,
nGraspSamples, graspScoreThresh, nGraspInliers, graspDetectMode)
rlAgent.PlotGrasps(grasps)
if len(grasps) == 0:
print("No grasps found. Skipping iteration.")
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotCloud()
continue
if showEveryStep:
raw_input("Press [Enter] to continue...")
# perform pick action
grasp = rlAgent.GetGrasp(grasps, epsilon)
rlAgent.PlotGrasps([grasp])
if showEveryStep:
raw_input("Press [Enter] to continue...")
# perform place action
P = rlAgent.GetPlacePose(grasp, epsilon)
rlAgent.MoveHandToPose(P)
rlAgent.MoveObjectToHandAtGrasp(grasp, objHandle)
r = rlEnv.RewardHeightExponential(
objHandle, targetObjectAxis, maxAngleFromObjectAxis, maxObjectTableGap)
print("The robot receives {} reward.".format(r))
Return.append(r)
if showEveryStep:
raw_input("Press [Enter] to continue...")
# cleanup this data iteration
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotGrasps()
rlAgent.UnplotCloud()
saveData = {"randomScale":randomScale, "targetObjectAxis":targetObjectAxis,
"maxAngleFromObjectAxis":maxAngleFromObjectAxis, "maxObjectTableGap":maxObjectTableGap,
"graspDetectMode":graspDetectMode, "nGraspSamples":nGraspSamples,
"graspScoreThresh":graspScoreThresh, "weightsFileName":weightsFileName,
"nDataIterations":nDataIterations, "epsilon":epsilon, "Return":Return}
savemat(saveFileName, saveData)
def Main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
params = loadmat("parameters.mat", squeeze_me=True)
randomSeed = params["randomSeed"]
tMax = params["tMax"]
nEpisodes = params["nEpisodes"]
trainEvery = params["trainEvery"]
unbiasOnEpisode = params["unbiasOnEpisode"]
saveFileName = params["saveFileName"]
loadNetwork = params["loadNetwork"]
loadDatabase = params["loadDatabase"]
showSteps = params["showSteps"]
# INITIALIZATION =================================================================================
# set random seeds
seed(randomSeed)
tensorflow.random.set_seed(randomSeed)
# initialize agent and environment
rlEnv = RlEnvironmentPegsOnDisks(params)
rlAgent = RlAgent(params)
# if testing, load previous results
if loadNetwork:
rlAgent.LoadQFunction()
if loadDatabase:
rlAgent.LoadExperienceDatabase()
# RUN TEST =======================================================================================
episodeReturn = []
nPlacedObjects = []
episodeTime = []
timeStepEpsilon = []
databaseSize = []
losses = []
for episode in xrange(nEpisodes):
startTime = time()
# place random object in random orientation on table
rlEnv.MoveHandToHoldingPose()
rlEnv.PlaceObjects(False)
rlEnv.PlaceObjects(True)
if showSteps: raw_input("Placed objects.")
R = 0
nPlaced = 0
holdingDesc = None
o = None
a = None
r = None
for t in xrange(tMax):
# get a point cloud
cloud = rlEnv.GetArtificialCloud()
isGrasp = holdingDesc is None
# get the next action
oo, aa, overtDesc, epsilon = rlAgent.SenseAndAct(
holdingDesc, cloud, t, rlEnv, episode >= unbiasOnEpisode)
# perform transition
holdingDesc, rr = rlEnv.Transition(overtDesc, cloud)
# save experience
if t > 0: rlAgent.AddExperienceSarsa(o, a, r, oo, aa)
o = oo
a = aa
r = rr
# save recorded data
R += r
if isGrasp:
if r < 0:
nPlaced -= 1
else:
if r > 0:
nPlaced += 1
timeStepEpsilon.append(epsilon)
# add final experience
rlAgent.AddExperienceSarsa(o, a, r, [None] * rlAgent.nLevels,
[None] * rlAgent.nLevels)
# cleanup episode
rlEnv.ResetEpisode()
print("Episode {} had return {}".format(episode, R))
# training
if episode % trainEvery == trainEvery - 1:
losses.append(rlAgent.UpdateQFunctionSarsa())
rlAgent.SaveQFunction()
# save results
episodeReturn.append(R)
nPlacedObjects.append(nPlaced)
episodeTime.append(time() - startTime)
databaseSize.append(rlAgent.GetNumberOfExperiences())
if episode % trainEvery == trainEvery - 1 or episode == nEpisodes - 1:
saveData = {
"episodeReturn": episodeReturn,
"nPlacedObjects": nPlacedObjects,
"episodeTime": episodeTime,
"timeStepEpsilon": timeStepEpsilon,
"databaseSize": databaseSize,
"losses": losses
}
saveData.update(params)
savemat(saveFileName, saveData)
# backup agent data
if episode == nEpisodes - 1:
rlAgent.SaveExperienceDatabase()
def main(objectClass, epsilon):
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# objects
nObjects = 7
randomObjectScale = True
targetObjectAxis = array([0, 0, 1])
maxAngleFromObjectAxis = 20 * (pi / 180)
maxObjectTableGap = 0.02
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.50
viewWorkspace = [(-1, 1), (-1, 1), (-1, 1)]
objViewWorkspace = [(-1, 1), (-1, 1), (0.002, 1)]
# grasps
graspDetectMode = 1 # 0=sample, 1=sample+label
nGraspSamples = 500
graspScoreThresh = 300
nGraspInliers = 2
# testing
weightsFileName = "/home/mgualti/mgualti/PickAndPlace/simulation/caffe/image_iter_5000.caffemodel"
nDataIterations = 300
# visualization/saving
saveFileName = "results-clutter-" + objectClass + "-epsilon" + str(
epsilon) + ".mat"
showViewer = False
showSteps = False
# INITIALIZATION =================================================================================
threeDNet = ThreeDNet()
rlEnv = RlEnvironment(showViewer)
rlAgent = RlAgent(rlEnv)
if epsilon < 1.0: rlAgent.LoadNetworkWeights(weightsFileName)
Return = []
# RUN TEST =======================================================================================
for dataIterationIdx in xrange(nDataIterations):
print("Iteration {}.".format(dataIterationIdx))
# place clutter on table
fullObjNames, objScales = threeDNet.GetRandomObjectSet(
objectClass, nObjects, randomObjectScale)
objHandles, objPoses = rlEnv.PlaceObjectSet(fullObjNames, objScales)
objCloud, objCloudIdxs = rlEnv.AssignPointsToObjects(
rlAgent, objHandles, viewCenter, viewKeepout, objViewWorkspace)
if showSteps:
raw_input("Objects placed.")
# move the hand to view position and capture a point cloud
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.PlotCloud(cloud)
if showSteps:
raw_input("Point cloud.")
# detect grasps in the sensory data
graspsDetected = rlAgent.DetectGrasps(cloud, viewPoints,
viewPointIndices, nGraspSamples,
graspScoreThresh, nGraspInliers,
graspDetectMode)
grasps = rlAgent.FilterGraspsWithNoPoints(graspsDetected, objCloud)
if len(graspsDetected) > len(grasps):
print("Fitlered {} empty grasps.".format(
len(graspsDetected) - len(grasps)))
rlAgent.PlotGrasps(grasps)
if showSteps:
raw_input("Acquired grasps.")
if len(grasps) == 0:
print("No grasps found. Skipping iteration.")
rlEnv.RemoveObjectSet(objHandles)
rlAgent.UnplotGrasps()
rlAgent.UnplotCloud()
continue
# perform pick action
grasp = rlAgent.GetGrasp(grasps, epsilon)
rlAgent.PlotGrasps([grasp])
if showSteps:
raw_input("Selected grasp.")
# perform place action
P = rlAgent.GetPlacePose(grasp, epsilon)
rlAgent.MoveHandToPose(P)
objHandle, objPose = rlEnv.GetObjectWithMaxGraspPoints(
grasp, objHandles, objCloud, objCloudIdxs)
rlAgent.MoveObjectToHandAtGrasp(grasp, objHandle)
r = rlEnv.RewardHeightExponential(objHandle, targetObjectAxis,
maxAngleFromObjectAxis,
maxObjectTableGap)
print("The robot receives {} reward.".format(r))
Return.append(r)
if showSteps:
raw_input("Press [Enter] to continue...")
# cleanup this data iteration
rlEnv.RemoveObjectSet(objHandles)
rlAgent.UnplotGrasps()
rlAgent.UnplotCloud()
# Save results
saveData = {
"nObjects": nObjects,
"randomObjectScale": randomObjectScale,
"targetObjectAxis": targetObjectAxis,
"maxAngleFromObjectAxis": maxAngleFromObjectAxis,
"maxObjectTableGap": maxObjectTableGap,
"graspDetectMode": graspDetectMode,
"nGraspSamples": nGraspSamples,
"graspScoreThresh": graspScoreThresh,
"weightsFileName": weightsFileName,
"nDataIterations": nDataIterations,
"epsilon": epsilon,
"Return": Return
}
savemat(saveFileName, saveData)
def main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# system
gpuId = 0
# objects
nObjects = 10
objectFolder = "/home/mgualti/Data/RaveObjects/RectangularBlocks"
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.70
viewWorkspace = [(-1.0, 1.0), (-1.0, 1.0), (-1.0, 1.0)]
# learning
nEpisodes = 1000
gpdScoreThresh = -float('Inf')
gpdNSamples = 500
# visualization/saving
saveFileName = "results.mat"
showViewer = False
showSteps = False
# INITIALIZATION =================================================================================
rlEnv = RlEnvironmentGrasping(showViewer)
rlAgent = RlAgent(rlEnv, gpuId)
gpd = GraspProxyMatlab()
# RUN TEST =======================================================================================
antipodal = []
perfect = []
episodeTime = []
for episode in xrange(nEpisodes):
# Initialization
episodeStartTime = time.time()
# place random object in random orientation on table
objHandles = rlEnv.PlaceObjects(nObjects, objectFolder)
if showSteps: raw_input("Placed objects.")
cloud, cloudTree, viewPoints, viewPointIndices = rlAgent.GetDualCloudAndViewPoints(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.PlotCloud(cloud)
if showSteps: raw_input("Acquired point cloud.")
#SaveCloud(cloud, viewPoints, viewPointIndices, "blocks1.mat")
# call gpd
descriptors = gpd.DetectGrasps(cloud, viewPoints, viewPointIndices,
gpdNSamples, gpdScoreThresh, gpuId)
if len(descriptors) > 0:
# choose descriptor with max score
bestScore = -float('inf')
for descriptor in descriptors:
if descriptor.score > bestScore:
desc = descriptor
bestScore = desc.score
rlAgent.PlotDescriptors([desc])
# check grasp and finish
ant, antAndCf = rlEnv.TestGrasp(desc, rlAgent, objHandles)
else:
print("No grasps found!")
ant = 0.0
perfect = 0.0
# cleanup this scene
print("Episode {}, antipodal={}, antipodal+collisionFree={}".format(
episode, ant, antAndCf))
antipodal.append(ant)
perfect.append(antAndCf)
rlEnv.RemoveObjectSet(objHandles)
# Save results
episodeTime.append(time.time() - episodeStartTime)
saveData = {
"gpuId": gpuId,
"nObjects": nObjects,
"objectFolder": objectFolder,
"viewCenter": viewCenter,
"viewKeepout": viewKeepout,
"viewWorkspace": viewWorkspace,
"nEpisodes": nEpisodes,
"gpdScoreThresh": gpdScoreThresh,
"gpdNSamples": gpdNSamples,
"episodeTime": episodeTime,
"antipodal": antipodal,
"perfect": perfect
}
savemat(saveFileName, saveData)
def main(objectClass):
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
# objects
randomScale = True
targetObjectAxis = array([0, 0, 1])
maxAngleFromObjectAxis = 20 * (pi / 180)
maxObjectTableGap = 0.03
# view
viewCenter = array([0, 0, 0])
viewKeepout = 0.50
viewWorkspace = [(-1, 1), (-1, 1), (-1, 1)]
# grasps
graspDetectMode = 1 # 0=sample, 1=sample+label
nGraspSamples = 200
graspScoreThresh = 300
# testing
nEpisodes = 300
maxTimesteps = 10
gamma = 0.98
epsilon = 0.0
weightsFileName = \
"/home/mgualti/mgualti/PickAndPlace/simulation/caffe/dualImage_iter_5000.caffemodel"
# visualization/saving
saveFileName = "results-" + objectClass + ".mat"
showViewer = False
showSteps = False
# INITIALIZATION =================================================================================
threeDNet = ThreeDNet()
rlEnv = RlEnvironment(showViewer)
rlAgent = RlAgent(rlEnv)
rlAgent.LoadNetworkWeights(weightsFileName)
nPlaceOptions = len(rlAgent.placePoses)
placeHistogram = zeros(nPlaceOptions)
Return = []
graspsDetected = []
topGraspsDetected = []
goodTempPlaceCount = []
badTempPlaceCount = []
goodFinalPlaceCount = []
badFinalPlaceCount = []
# RUN TEST =======================================================================================
# for each episode/object placement
for episode in xrange(nEpisodes):
# place random object in random orientation on table
fullObjName, objScale = threeDNet.GetRandomObjectFromClass(
objectClass, randomScale)
objHandle, objRandPose = rlEnv.PlaceObjectRandomOrientation(
fullObjName, objScale)
rlAgent.MoveSensorToPose(
rlAgent.GetStandardViewPose(viewCenter, viewKeepout))
if showSteps:
raw_input("Beginning of episode. Press [Enter] to continue...")
# move the hand to view position(s) and capture a point cloud
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.PlotCloud(cloud)
if showSteps:
raw_input("Acquired point cloud. Press [Enter] to continue...")
# detect grasps in the sensor data
grasps = rlAgent.DetectGrasps(cloud, viewPoints, viewPointIndices,
nGraspSamples, graspScoreThresh,
graspDetectMode)
graspsDetected.append(len(grasps))
topGraspsCount = CountObjectTopGrasps(grasps, objRandPose,
maxAngleFromObjectAxis)
if len(grasps) == 0:
print("No grasps found. Skipping iteration.")
rlEnv.RemoveObject(objHandle)
continue
rlAgent.PlotGrasps(grasps)
print("Episode {}.".format(episode))
if showSteps:
raw_input("Acquired grasps. Press [Enter] to continue...")
# initialize recording variables
episodePlaceHistogram = zeros(nPlaceOptions)
episodeReturn = 0
episodeGoodTempPlaceCount = 0
episodeBadTempPlaceCount = 0
episodeGoodFinalPlaceCount = 0
episodeBadFinalPlaceCount = 0
graspDetectionFailure = False
# initial state and first action
s, selectedGrasp = rlEnv.GetInitialState(rlAgent)
a, grasp, place = rlAgent.ChooseAction(s, grasps, epsilon)
rlAgent.PlotGrasps([grasp])
# for each time step in the episode
for t in xrange(maxTimesteps):
ss, selectedGrasp, rr = rlEnv.Transition(rlAgent, objHandle, s,
selectedGrasp, a, grasp,
place, targetObjectAxis,
maxAngleFromObjectAxis,
maxObjectTableGap)
ssIsPlacedTempGood = ss[1][1]
ssIsPlacedTempBad = ss[1][2]
ssIsPlacedFinalGood = ss[1][3]
ssIsPlacedFinalBad = ss[1][4]
if showSteps:
raw_input(
"Transition {}. Press [Enter] to continue...".format(t))
# re-detect only if a non-terminal placement just happened
if ssIsPlacedTempGood and place is not None:
cloud, viewPoints, viewPointIndices = rlAgent.GetDualCloud(
viewCenter, viewKeepout, viewWorkspace)
rlAgent.UnplotGrasps()
rlAgent.PlotCloud(cloud)
if showSteps:
raw_input("Acquired cloud. Press [Enter] to continue...")
grasps = rlAgent.DetectGrasps(cloud, viewPoints,
viewPointIndices, nGraspSamples,
graspScoreThresh,
graspDetectMode)
graspsDetected.append(len(grasps))
topGraspsCount = CountObjectTopGrasps(grasps, objRandPose,
maxAngleFromObjectAxis)
topGraspsDetected.append(topGraspsCount)
if len(grasps) == 0:
print("Grasp detection failure.")
graspDetectionFailure = True
break
rlAgent.PlotGrasps(grasps)
if showSteps:
raw_input("Acquired grasps. Press [Enter] to continue...")
# get next action
aa, ggrasp, pplace = rlAgent.ChooseAction(ss, grasps, epsilon)
if ggrasp is not None: rlAgent.PlotGrasps([ggrasp])
if showSteps:
raw_input("Action {}. Press [Enter] to continue...".format(t))
# record data from transition
episodeReturn += (gamma**t) * rr
if place is not None:
episodeGoodTempPlaceCount += ssIsPlacedTempGood
episodeBadTempPlaceCount += ssIsPlacedTempBad
episodeGoodFinalPlaceCount += ssIsPlacedFinalGood
episodeBadFinalPlaceCount += ssIsPlacedFinalBad
placeHistogram += a[1]
# prepare for next time step
if ssIsPlacedTempBad or ssIsPlacedFinalGood or ssIsPlacedFinalBad:
break
s = ss
a = aa
grasp = ggrasp
place = pplace
# cleanup this reuse
if not graspDetectionFailure:
placeHistogram += episodePlaceHistogram
Return.append(episodeReturn)
goodTempPlaceCount.append(episodeGoodTempPlaceCount)
badTempPlaceCount.append(episodeBadTempPlaceCount)
goodFinalPlaceCount.append(episodeGoodFinalPlaceCount)
badFinalPlaceCount.append(episodeBadFinalPlaceCount)
# cleanup this episode
rlEnv.RemoveObject(objHandle)
rlAgent.UnplotGrasps()
rlAgent.UnplotCloud()
# Save results
saveData = {
"objectClass": objectClass,
"randomScale": randomScale,
"maxAngleFromObjectAxis": maxAngleFromObjectAxis,
"maxObjectTableGap": maxObjectTableGap,
"nGraspSamples": nGraspSamples,
"graspScoreThresh": graspScoreThresh,
"graspDetectMode": graspDetectMode,
"nEpisodes": nEpisodes,
"maxTimesteps": maxTimesteps,
"gamma": gamma,
"epsilon": epsilon,
"Return": Return,
"graspsDetected": graspsDetected,
"topGraspsDetected": topGraspsDetected,
"placeHistogram": placeHistogram,
"goodTempPlaceCount": goodTempPlaceCount,
"badTempPlaceCount": badTempPlaceCount,
"goodFinalPlaceCount": goodFinalPlaceCount,
"badFinalPlaceCount": badFinalPlaceCount
}
savemat(saveFileName, saveData)
def Main():
'''Entrypoint to the program.'''
# PARAMETERS =====================================================================================
params = loadmat("parameters.mat", squeeze_me=True)
randomSeed = params["randomSeed"]
tMax = params["tMax"]
nEpisodes = params["nEpisodes"]
trainEvery = params["trainEvery"]
unbiasOnEpisode = params["unbiasOnEpisode"]
nObjects = params["nObjects"]
nSurfaceObjects = params["nSurfaceObjects"]
objHeight = params["objHeight"]
objRadius = params["objRadius"]
surfObjHeight = params["surfObjHeight"]
surfObjRadius = params["surfObjRadius"]
saveFileName = params["saveFileName"]
loadNetwork = params["loadNetwork"]
loadDatabase = params["loadDatabase"]
showSteps = params["showSteps"]
# INITIALIZATION =================================================================================
# set random seeds
seed(randomSeed)
tensorflow.random.set_seed(randomSeed)
# initialize agent and environment
rlEnv = RlEnvironmentPegsOnDisks(params)
rlAgent = RlAgent(params)
# if testing, load previous results
if loadNetwork:
rlAgent.LoadQFunction()
if loadDatabase:
rlAgent.LoadExperienceDatabase()
# RUN TEST =======================================================================================
episodeReturn = []
episodeTime = []
timeStepEpsilon = []
databaseSize = []
losses = []
for episode in xrange(nEpisodes):
startTime = time()
# place random object in random orientation on table
rlEnv.MoveHandToHoldingPose()
rlEnv.PlaceCylinders(nObjects, objHeight, objRadius, False)
rlEnv.PlaceCylinders(nSurfaceObjects, surfObjHeight, surfObjRadius,
True)
if showSteps: raw_input("Placed objects.")
R = 0
holdingDesc = None
observations = []
actions = []
rewards = []
for t in xrange(tMax):
# get the next action
o, a, overtDesc, epsilon = rlAgent.SenseAndAct(
holdingDesc, t, rlEnv, episode >= unbiasOnEpisode)
# perform transition
holdingDesc, r = rlEnv.Transition(overtDesc)
# save experiences
timeStepEpsilon.append(epsilon)
observations.append(o)
actions.append(a)
rewards.append(r)
R += r
rlAgent.AddExperienceMonteCarlo(observations, actions, rewards)
# cleanup episode
rlEnv.ResetEpisode()
print("Episode {} had return {}".format(episode, R))
# training
if episode % trainEvery == trainEvery - 1:
losses.append(rlAgent.UpdateQFunctionMonteCarlo())
rlAgent.SaveQFunction()
# save results
episodeReturn.append(R)
episodeTime.append(time() - startTime)
databaseSize.append(rlAgent.GetNumberOfExperiences())
if episode % trainEvery == trainEvery - 1 or episode == nEpisodes - 1:
saveData = {
"episodeReturn": episodeReturn,
"episodeTime": episodeTime,
"timeStepEpsilon": timeStepEpsilon,
"databaseSize": databaseSize,
"losses": losses
}
saveData.update(params)
savemat(saveFileName, saveData)
# backup agent data
if episode == nEpisodes - 1:
rlAgent.SaveExperienceDatabase()