def OneTraj(self, x, y):
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1, q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
self.vectarget = [0.0, 0.0, self.rs.XTarget, self.rs.YTarget]
totalCost = 0
while (not self.isFinished()):
action = self.actor.action(self.estimState)
_, cost = self.__act__(action)
totalCost += cost[0]
totalCost += self.trajCost.computeFinalReward(self.arm, self.t,
self.coordHand,
self.sizeOfTarget)
return totalCost, self.t
def saveOneTraj(self, x, y):
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1,q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
self.dataStore=[]
self.vectarget=[0.0, 0.0, self.rs.XTarget, self.rs.YTarget]
totalCost=0
while(not self.isFinished()):
action=self.actor.action(self.estimState)
_,cost = self.actAndStore(action)
totalCost+= cost[0]
totalCost += self.trajCost.computeFinalReward(self.arm,self.t,self.coordHand,self.sizeOfTarget)
filename = findDataFilename(self.saveName+"Log/","traj"+str(x)+"-"+str(y),".log")
np.savetxt(filename,self.dataStore)
#used to ignore dispersion when the target line is not crossed
if self.coordHand[1] >= self.rs.YTarget:
self.lastCoord.append(self.coordHand[0])
return totalCost, self.t
def OneTraj(self, x, y):
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1,q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
self.vectarget=[0.0, 0.0, self.rs.XTarget, self.rs.YTarget]
totalCost=0
while(not self.isFinished()):
action=self.actor.action(self.estimState)
_,cost = self.__act__(action)
totalCost+= cost[0]
totalCost += self.trajCost.computeFinalReward(self.arm,self.t,self.coordHand,self.sizeOfTarget)
return totalCost, self.t
def reset(self, noise=True):
#print("Episode : "+str(self.nbReset)+ ", Progression :"+str(self.progress))
#self.nbReset+=1
#print(self.cost)
#Discrete begining
"""
i = rd.randint(0,len(self.posIni)-1)
#i=0
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(self.posIni[i][0], self.posIni[i][1])
"""
#Progress
"""
i = ((rd.random()*6*np.pi - 3*np.pi)*self.progressTab[self.progress]-6*np.pi)/12
j= (rd.random()*0.3-0.15)*self.progressTab[self.progress]+0.25
#i=0
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(self.rs.XTarget+ j*np.cos(i), self.rs.YTarget+ j*np.sin(i))
"""
"""
i = - 6*np.pi/12
j=rd.randint(1,4)/10.
#j=0.1
q1, q2 = self.arm.mgi(self.rs.XTarget+ j*np.cos(i), self.rs.YTarget+ j*np.sin(i))
"""
x, y = self.begin()
q1, q2 = self.arm.mgi(x,y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1,q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0., 0.
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
def reset(self, noise=True):
#print("Episode : "+str(self.nbReset)+ ", Progression :"+str(self.progress))
#self.nbReset+=1
#print(self.cost)
#Discrete begining
"""
i = rd.randint(0,len(self.posIni)-1)
#i=0
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(self.posIni[i][0], self.posIni[i][1])
"""
#Progress
"""
i = ((rd.random()*6*np.pi - 3*np.pi)*self.progressTab[self.progress]-6*np.pi)/12
j= (rd.random()*0.3-0.15)*self.progressTab[self.progress]+0.25
#i=0
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(self.rs.XTarget+ j*np.cos(i), self.rs.YTarget+ j*np.sin(i))
"""
"""
i = - 6*np.pi/12
j=rd.randint(1,4)/10.
#j=0.1
q1, q2 = self.arm.mgi(self.rs.XTarget+ j*np.cos(i), self.rs.YTarget+ j*np.sin(i))
"""
x, y = self.begin()
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1, q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0., 0.
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
def saveOneTraj(self, x, y):
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1, q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
self.coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
self.i, self.t, self.cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
self.estimState = state
self.dataStore = []
self.vectarget = [0.0, 0.0, self.rs.XTarget, self.rs.YTarget]
totalCost = 0
while (not self.isFinished()):
action = self.actor.action(self.estimState)
_, cost = self.actAndStore(action)
totalCost += cost[0]
totalCost += self.trajCost.computeFinalReward(self.arm, self.t,
self.coordHand,
self.sizeOfTarget)
filename = findDataFilename(self.saveName + "Log/",
"traj" + str(x) + "-" + str(y), ".log")
np.savetxt(filename, self.dataStore)
#used to ignore dispersion when the target line is not crossed
if self.coordHand[1] >= self.rs.YTarget:
self.lastCoord.append(self.coordHand[0])
return totalCost, self.t
def runTrajectory(self, x, y, foldername):
'''
Generates trajectory from the initial position (x, y)
Inputs: -x: abscissa of the initial position, float
-y: ordinate of the initial position, float
Output: -cost: the cost of the trajectory, float
'''
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1,q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
coordElbow, coordHand = self.arm.mgdFull(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
i, t, cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
estimState = state
dataStore = []
qtarget1, qtarget2 = self.arm.mgi(self.rs.XTarget, self.rs.YTarget)
vectarget = [0.0, 0.0, qtarget1, qtarget2]
#loop to generate next position until the target is reached
while coordHand[1] < self.rs.YTarget and i < self.rs.numMaxIter:
stepStore = []
#computation of the next muscular activation vector using the controller theta
#print ("state :",self.arm.getState())
U = self.controller.computeOutput(estimState)
if det:
Unoisy = muscleFilter(U)
else:
Unoisy = getNoisyCommand(U,self.arm.musclesP.knoiseU)
Unoisy = muscleFilter(Unoisy)
#computation of the arm state
realNextState = self.arm.computeNextState(Unoisy, self.arm.getState())
#computation of the approximated state
tmpState = self.arm.getState()
if det:
estimNextState = realNextState
else:
U = muscleFilter(U)
estimNextState = self.stateEstimator.getEstimState(tmpState,U)
#print estimNextState
self.arm.setState(realNextState)
#computation of the cost
cost += self.computeStateTransitionCost(Unoisy)
'''
print "U =", U
print "Unoisy =", Unoisy
print "estimstate =", estimState
#print "theta =", self.controller.theta
if math.isnan(cost):
print "NAN : U =", U
print "NAN : Unoisy =", Unoisy
print "NAN : estimstate =", estimState
#print "NAN : theta =", self.controller.theta
sys.exit()
'''
#get dotq and q from the state vector
dotq, q = getDotQAndQFromStateVector(tmpState)
coordElbow, coordHand = self.arm.mgdFull(q)
#print ("dotq :",dotq)
#computation of the coordinates to check if the target is reach or not
#code to save data of the trajectory
#Note : these structures might be much improved
if self.saveTraj == True:
stepStore.append(vectarget)
stepStore.append(estimState)
stepStore.append(tmpState)
stepStore.append(Unoisy)
stepStore.append(np.array(U))
stepStore.append(estimNextState)
stepStore.append(realNextState)
stepStore.append([coordElbow[0], coordElbow[1]])
stepStore.append([coordHand[0], coordHand[1]])
#print ("before",stepStore)
tmpstore = np.array(stepStore).flatten()
row = [item for sub in tmpstore for item in sub]
#print ("store",row)
dataStore.append(row)
estimState = estimNextState
i += 1
t += self.rs.dt
cost += self.computeFinalReward(t,coordHand)
if self.saveTraj == True:
filename = findFilename(foldername+"Log/","traj",".log")
np.savetxt(filename,dataStore)
'''
if coordHand[0] >= -self.sizeOfTarget/2 and coordHand[0] <= self.sizeOfTarget/2 and coordHand[1] >= self.rs.YTarget and i<230:
foldername = pathDataFolder + "TrajRepository/"
name = findFilename(foldername,"Traj",".traj")
np.savetxt(name,dataStore)
'''
lastX = -1000 #used to ignore dispersion when the target line is not crossed
if coordHand[1] >= self.rs.YTarget:
lastX = coordHand[0]
#print "end of trajectory"
return cost, t, lastX
def runTrajectoryOpti(self, x, y):
'''
Generates trajectory from the initial position (x, y) use for plot trajectory wihtout save them
Inputs: -x: abscissa of the initial position, float
-y: ordinate of the initial position, float
Output:
-cost: the cost of the trajectory, float
-t: time of the trajectory, float
-lastX: Last X
'''
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1,q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
i, t, cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
estimState = state
#loop to generate next position until the target is reached
while coordHand[1] < self.rs.YTarget and i < self.rs.maxSteps:
#computation of the next muscular activation vector using the controller theta
#print ("state :",self.arm.getState())
U = self.controller.computeOutput(estimState)
if self.rs.det:
realU = muscleFilter(U)
#computation of the arm state
realNextState = self.arm.computeNextState(realU, self.arm.getState())
#computation of the approximated state
tmpState = self.arm.getState()
estimNextState = realNextState
else:
#realU = getNoisyCommand(U,self.arm.getMusclesParameters().getKnoiseU())
realU = getNoisyCommand(U,self.arm.musclesP.knoiseU)
realU = muscleFilter(realU)
#computation of the arm state
realNextState = self.arm.computeNextState(realU, self.arm.getState())
#computation of the approximated state
tmpState = self.arm.getState()
U = muscleFilter(U)
estimNextState = self.stateEstimator.getEstimState(tmpState,U)
#print estimNextState
self.arm.setState(realNextState)
#computation of the cost
cost += self.trajCost.computeStateTransitionCost(realU)
#get dotq and q from the state vector
_, q = self.arm.getDotQAndQFromStateVector(tmpState)
coordHand = self.arm.mgdEndEffector(q)
#print ("dotq :",dotq)
#computation of the coordinates to check if the target is reach or not
estimState = estimNextState
i += 1
t += self.rs.dt
cost += self.trajCost.computeFinalReward(self.arm,t,coordHand, self.sizeOfTarget)
lastX = -1000 #used to ignore dispersion when the target line is not crossed
if coordHand[1] >= self.rs.YTarget:
lastX = coordHand[0]
return cost, t, lastX
def runTrajectoryOpti(self, x, y):
'''
Generates trajectory from the initial position (x, y) use for plot trajectory wihtout save them
Inputs: -x: abscissa of the initial position, float
-y: ordinate of the initial position, float
Output:
-cost: the cost of the trajectory, float
-t: time of the trajectory, float
-lastX: Last X
'''
#computes the articular position q1, q2 from the initial coordinates (x, y)
q1, q2 = self.arm.mgi(x, y)
#creates the state vector [dotq1, dotq2, q1, q2]
q = createVector(q1, q2)
state = np.array([0., 0., q1, q2])
#print("start state --------------: ",state)
#computes the coordinates of the hand and the elbow from the position vector
coordHand = self.arm.mgdEndEffector(q)
#assert(coordHand[0]==x and coordHand[1]==y), "Erreur de MGD" does not work because of rounding effects
#initializes parameters for the trajectory
i, t, cost = 0, 0, 0
self.stateEstimator.initStore(state)
self.arm.setState(state)
estimState = state
#loop to generate next position until the target is reached
while coordHand[1] < self.rs.YTarget and i < self.rs.maxSteps:
#computation of the next muscular activation vector using the controller theta
#print ("state :",self.arm.getState())
U = self.controller.computeOutput(estimState)
if self.rs.det:
realU = muscleFilter(U)
#computation of the arm state
realNextState = self.arm.computeNextState(
realU, self.arm.getState())
#computation of the approximated state
tmpState = self.arm.getState()
estimNextState = realNextState
else:
#realU = getNoisyCommand(U,self.arm.getMusclesParameters().getKnoiseU())
realU = getNoisyCommand(U, self.arm.musclesP.knoiseU)
realU = muscleFilter(realU)
#computation of the arm state
realNextState = self.arm.computeNextState(
realU, self.arm.getState())
#computation of the approximated state
tmpState = self.arm.getState()
U = muscleFilter(U)
estimNextState = self.stateEstimator.getEstimState(tmpState, U)
#print estimNextState
self.arm.setState(realNextState)
#computation of the cost
cost += self.trajCost.computeStateTransitionCost(realU)
#get dotq and q from the state vector
_, q = self.arm.getDotQAndQFromStateVector(tmpState)
coordHand = self.arm.mgdEndEffector(q)
#print ("dotq :",dotq)
#computation of the coordinates to check if the target is reach or not
estimState = estimNextState
i += 1
t += self.rs.dt
cost += self.trajCost.computeFinalReward(self.arm, t, coordHand,
self.sizeOfTarget)
lastX = -1000 #used to ignore dispersion when the target line is not crossed
if coordHand[1] >= self.rs.YTarget:
lastX = coordHand[0]
return cost, t, lastX