class FramePlacementCostSumTest(CostModelSumTestCase):
ROBOT_MODEL = pinocchio.buildSampleModelHumanoidRandom()
ROBOT_STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
Mref = crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId('rleg5_joint'),
pinocchio.SE3.Random())
COST = crocoddyl.CostModelFramePlacement(ROBOT_STATE, Mref)
def createPseudoImpulseModel(self, supportFootIds, swingFootTask):
""" Action model for pseudo-impulse models.
A pseudo-impulse model consists of adding high-penalty cost for the contact velocities.
:param swingFootTask: swinging foot task
:return pseudo-impulse differential action model
"""
# Creating a 6D multi-contact model, and then including the supporting
# foot
contactModel = crocoddyl.ContactModelMultiple(self.state,
self.actuation.nu)
for i in supportFootIds:
Mref = crocoddyl.FramePlacement(i, pinocchio.SE3.Identity())
supportContactModel = crocoddyl.ContactModel6D(
self.state, Mref, self.actuation.nu, np.array([0., 0.]))
contactModel.addContact(self.rmodel.frames[i].name + "_contact",
supportContactModel)
# Creating the cost model for a contact phase
costModel = crocoddyl.CostModelSum(self.state, self.actuation.nu)
for i in supportFootIds:
cone = crocoddyl.WrenchCone(np.identity(3), self.mu,
np.array([0.1, 0.05]))
wrenchCone = crocoddyl.CostModelContactWrenchCone(
self.state,
crocoddyl.ActivationModelQuadraticBarrier(
crocoddyl.ActivationBounds(cone.lb, cone.ub)),
crocoddyl.FrameWrenchCone(i, cone), self.actuation.nu)
costModel.addCost(self.rmodel.frames[i].name + "_wrenchCone",
wrenchCone, 1e1)
if swingFootTask is not None:
for i in swingFootTask:
footTrack = crocoddyl.CostModelFramePlacement(
self.state, i, self.actuation.nu)
costModel.addCost(self.rmodel.frames[i.id].name + "_footTrack",
footTrack, 1e8)
footVel = crocoddyl.FrameMotion(i.id, pinocchio.Motion.Zero())
impulseFootVelCost = crocoddyl.CostModelFrameVelocity(
self.state, footVel, self.actuation.nu)
costModel.addCost(
self.rmodel.frames[i.id].name + "_impulseVel",
impulseFootVelCost, 1e6)
stateWeights = np.array([0] * 3 + [500.] * 3 + [0.01] *
(self.state.nv - 6) + [10] * self.state.nv)
stateReg = crocoddyl.CostModelState(
self.state, crocoddyl.ActivationModelWeightedQuad(stateWeights**2),
self.rmodel.defaultState, self.actuation.nu)
ctrlReg = crocoddyl.CostModelControl(self.state, self.actuation.nu)
costModel.addCost("stateReg", stateReg, 1e1)
costModel.addCost("ctrlReg", ctrlReg, 1e-3)
# Creating the action model for the KKT dynamics with simpletic Euler
# integration scheme
dmodel = crocoddyl.DifferentialActionModelContactFwdDynamics(
self.state, self.actuation, contactModel, costModel, 0., True)
model = crocoddyl.IntegratedActionModelEuler(dmodel, 0.)
return model
class FramePlacementCostSumTest(CostModelSumTestCase):
ROBOT_MODEL = example_robot_data.load('icub_reduced').model
ROBOT_STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
Mref = crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId('r_sole'),
pinocchio.SE3.Random())
COST = crocoddyl.CostModelFramePlacement(ROBOT_STATE, Mref)
class FramePlacementCostTest(CostModelAbstractTestCase):
ROBOT_MODEL = example_robot_data.loadICub().model
ROBOT_STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
Mref = crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId('r_sole'),
pinocchio.SE3.Random())
COST = crocoddyl.CostModelFramePlacement(ROBOT_STATE, Mref)
COST_DER = FramePlacementCostModelDerived(ROBOT_STATE, Mref=Mref)
def runBenchmark(model):
robot_model = ROBOT.model
q0 = np.matrix([0.173046, 1., -0.52366, 0., 0., 0.1, -0.005]).T
x0 = np.vstack([q0, np.zeros((robot_model.nv, 1))])
# Note that we need to include a cost model (i.e. set of cost functions) in
# order to fully define the action model for our optimal control problem.
# For this particular example, we formulate three running-cost functions:
# goal-tracking cost, state and control regularization; and one terminal-cost:
# goal cost. First, let's create the common cost functions.
state = crocoddyl.StateMultibody(robot_model)
Mref = crocoddyl.FramePlacement(
robot_model.getFrameId("gripper_left_joint"),
pinocchio.SE3(np.eye(3), np.matrix([[.0], [.0], [.4]])))
goalTrackingCost = crocoddyl.CostModelFramePlacement(state, Mref)
xRegCost = crocoddyl.CostModelState(state)
uRegCost = crocoddyl.CostModelControl(state)
# Create a cost model per the running and terminal action model.
runningCostModel = crocoddyl.CostModelSum(state)
terminalCostModel = crocoddyl.CostModelSum(state)
# Then let's added the running and terminal cost functions
runningCostModel.addCost("gripperPose", goalTrackingCost, 1e-3)
runningCostModel.addCost("xReg", xRegCost, 1e-7)
runningCostModel.addCost("uReg", uRegCost, 1e-7)
terminalCostModel.addCost("gripperPose", goalTrackingCost, 1)
# Next, we need to create an action model for running and terminal knots. The
# forward dynamics (computed using ABA) are implemented
# inside DifferentialActionModelFullyActuated.
runningModel = crocoddyl.IntegratedActionModelEuler(
model(state, runningCostModel), 1e-3)
runningModel.differential.armature = np.matrix(
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.]).T
terminalModel = crocoddyl.IntegratedActionModelEuler(
model(state, terminalCostModel), 1e-3)
terminalModel.differential.armature = np.matrix(
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.]).T
# For this optimal control problem, we define 100 knots (or running action
# models) plus a terminal knot
problem = crocoddyl.ShootingProblem(x0, [runningModel] * N, terminalModel)
# Creating the DDP solver for this OC problem, defining a logger
ddp = crocoddyl.SolverDDP(problem)
duration = []
for i in range(T):
c_start = time.time()
ddp.solve([], [], MAXITER)
c_end = time.time()
duration.append(1e3 * (c_end - c_start))
avrg_duration = sum(duration) / len(duration)
min_duration = min(duration)
max_duration = max(duration)
return avrg_duration, min_duration, max_duration
def createSwingFootModel(self,
timeStep,
supportFootIds,
comTask=None,
swingFootTask=None):
""" Action model for a swing foot phase.
:param timeStep: step duration of the action model
:param supportFootIds: Ids of the constrained feet
:param comTask: CoM task
:param swingFootTask: swinging foot task
:return action model for a swing foot phase
"""
# Creating a 6D multi-contact model, and then including the supporting
# foot
contactModel = crocoddyl.ContactModelMultiple(self.state,
self.actuation.nu)
for i in supportFootIds:
Mref = crocoddyl.FramePlacement(i, pinocchio.SE3.Identity())
supportContactModel = \
crocoddyl.ContactModel6D(self.state, Mref, self.actuation.nu, np.matrix([0., 0.]).T)
contactModel.addContact(self.rmodel.frames[i].name + "_contact",
supportContactModel)
# Creating the cost model for a contact phase
costModel = crocoddyl.CostModelSum(self.state, self.actuation.nu)
if isinstance(comTask, np.ndarray):
comTrack = crocoddyl.CostModelCoMPosition(self.state, comTask,
self.actuation.nu)
costModel.addCost("comTrack", comTrack, 1e6)
if swingFootTask is not None:
for i in swingFootTask:
footTrack = crocoddyl.CostModelFramePlacement(
self.state, i, self.actuation.nu)
costModel.addCost(
self.rmodel.frames[i.frame].name + "_footTrack", footTrack,
1e6)
stateWeights = np.array([0] * 3 + [500.] * 3 + [0.01] *
(self.state.nv - 6) + [10] * self.state.nv)
stateReg = crocoddyl.CostModelState(
self.state,
crocoddyl.ActivationModelWeightedQuad(
np.matrix(stateWeights**2).T), self.rmodel.defaultState,
self.actuation.nu)
ctrlReg = crocoddyl.CostModelControl(self.state, self.actuation.nu)
costModel.addCost("stateReg", stateReg, 1e1)
costModel.addCost("ctrlReg", ctrlReg, 1e-1)
# Creating the action model for the KKT dynamics with simpletic Euler
# integration scheme
dmodel = crocoddyl.DifferentialActionModelContactFwdDynamics(
self.state, self.actuation, contactModel, costModel)
model = crocoddyl.IntegratedActionModelEuler(dmodel, timeStep)
return model
class FreeFwdDynamicsTest(ActionModelAbstractTestCase):
ROBOT_MODEL = pinocchio.buildSampleModelManipulator()
STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
COST_SUM = crocoddyl.CostModelSum(STATE, ROBOT_MODEL.nv)
COST_SUM.addCost('xReg', crocoddyl.CostModelState(STATE), 1.)
COST_SUM.addCost(
'frTrack',
crocoddyl.CostModelFramePlacement(
STATE,
crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId("effector_body"),
pinocchio.SE3.Random())), 1.)
MODEL = crocoddyl.DifferentialActionModelFreeFwdDynamics(STATE, COST_SUM)
MODEL_DER = DifferentialFreeFwdDynamicsDerived(STATE, COST_SUM)
def createProblem(model):
robot_model = ROBOT.model
q0 = np.matrix([0.173046, 1., -0.52366, 0., 0., 0.1, -0.005]).T
x0 = np.vstack([q0, np.zeros((robot_model.nv, 1))])
# Note that we need to include a cost model (i.e. set of cost functions) in
# order to fully define the action model for our optimal control problem.
# For this particular example, we formulate three running-cost functions:
# goal-tracking cost, state and control regularization; and one terminal-cost:
# goal cost. First, let's create the common cost functions.
state = crocoddyl.StateMultibody(robot_model)
Mref = crocoddyl.FramePlacement(
robot_model.getFrameId("gripper_left_joint"),
pinocchio.SE3(np.eye(3), np.matrix([[.0], [.0], [.4]])))
goalTrackingCost = crocoddyl.CostModelFramePlacement(state, Mref)
xRegCost = crocoddyl.CostModelState(state)
uRegCost = crocoddyl.CostModelControl(state)
# Create a cost model per the running and terminal action model.
runningCostModel = crocoddyl.CostModelSum(state)
terminalCostModel = crocoddyl.CostModelSum(state)
# Then let's added the running and terminal cost functions
runningCostModel.addCost("gripperPose", goalTrackingCost, 1)
runningCostModel.addCost("xReg", xRegCost, 1e-4)
runningCostModel.addCost("uReg", uRegCost, 1e-4)
terminalCostModel.addCost("gripperPose", goalTrackingCost, 1)
# Next, we need to create an action model for running and terminal knots. The
# forward dynamics (computed using ABA) are implemented
# inside DifferentialActionModelFullyActuated.
actuation = crocoddyl.ActuationModelFull(state)
runningModel = crocoddyl.IntegratedActionModelEuler(
model(state, actuation, runningCostModel), 1e-3)
runningModel.differential.armature = np.matrix(
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.]).T
terminalModel = crocoddyl.IntegratedActionModelEuler(
model(state, actuation, terminalCostModel), 1e-3)
terminalModel.differential.armature = np.matrix(
[0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.]).T
# For this optimal control problem, we define 100 knots (or running action
# models) plus a terminal knot
problem = crocoddyl.ShootingProblem(x0, [runningModel] * N, terminalModel)
xs = [x0] * (len(problem.runningModels) + 1)
us = [
m.quasiStatic(d, x0)
for m, d in list(zip(problem.runningModels, problem.runningDatas))
]
return xs, us, problem
class TalosArmFreeFwdDynamicsTest(ActionModelAbstractTestCase):
ROBOT_MODEL = example_robot_data.loadTalosArm().model
STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
ACTUATION = crocoddyl.ActuationModelFull(STATE)
COST_SUM = crocoddyl.CostModelSum(STATE)
COST_SUM.addCost(
'gripperPose',
crocoddyl.CostModelFramePlacement(
STATE, crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId("gripper_left_joint"), pinocchio.SE3.Random())),
1e-3)
COST_SUM.addCost("xReg", crocoddyl.CostModelState(STATE), 1e-7)
COST_SUM.addCost("uReg", crocoddyl.CostModelControl(STATE), 1e-7)
MODEL = crocoddyl.DifferentialActionModelFreeFwdDynamics(STATE, ACTUATION, COST_SUM)
MODEL_DER = DifferentialFreeFwdDynamicsDerived(STATE, ACTUATION, COST_SUM)
class ManipulatorDDPTest(SolverAbstractTestCase):
ROBOT_MODEL = pinocchio.buildSampleModelManipulator()
STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
COST_SUM = crocoddyl.CostModelSum(STATE, ROBOT_MODEL.nv)
COST_SUM.addCost('xReg', crocoddyl.CostModelState(STATE), 1e-7)
COST_SUM.addCost('uReg', crocoddyl.CostModelControl(STATE), 1e-7)
COST_SUM.addCost(
'frTrack',
crocoddyl.CostModelFramePlacement(
STATE, crocoddyl.FramePlacement(ROBOT_MODEL.getFrameId("effector_body"), pinocchio.SE3.Random())), 1.)
DIFF_MODEL = crocoddyl.DifferentialActionModelFreeFwdDynamics(STATE, COST_SUM)
MODEL = crocoddyl.IntegratedActionModelEuler(crocoddyl.DifferentialActionModelFreeFwdDynamics(STATE, COST_SUM),
1e-3)
SOLVER = crocoddyl.SolverDDP
SOLVER_DER = DDPDerived
class TalosArmFDDPTest(SolverAbstractTestCase):
ROBOT_MODEL = example_robot_data.load('talos_arm').model
STATE = crocoddyl.StateMultibody(ROBOT_MODEL)
ACTUATION = crocoddyl.ActuationModelFull(STATE)
COST_SUM = crocoddyl.CostModelSum(STATE)
COST_SUM.addCost(
'gripperPose',
crocoddyl.CostModelFramePlacement(
STATE,
crocoddyl.FramePlacement(
ROBOT_MODEL.getFrameId("gripper_left_joint"),
pinocchio.SE3.Random())), 1e-3)
COST_SUM.addCost("xReg", crocoddyl.CostModelState(STATE), 1e-7)
COST_SUM.addCost("uReg", crocoddyl.CostModelControl(STATE), 1e-7)
DIFF_MODEL = crocoddyl.DifferentialActionModelFreeFwdDynamics(
STATE, ACTUATION, COST_SUM)
MODEL = crocoddyl.IntegratedActionModelEuler(DIFF_MODEL, 1e-3)
SOLVER = crocoddyl.SolverFDDP
SOLVER_DER = FDDPDerived
def createFootSwitchModel(self, supportFootIds, swingFootTask):
""" Action model for a foot switch phase.
:param supportFootIds: Ids of the constrained feet
:param swingFootTask: swinging foot task
:return action model for a foot switch phase
"""
# Creating a 6D multi-contact model, and then including the supporting
# foot
contactModel = crocoddyl.ContactModelMultiple(self.state, self.actuation.nu)
for i in supportFootIds:
Mref = crocoddyl.FramePlacement(i, pinocchio.SE3.Identity())
supportContactModel = crocoddyl.ContactModel6D(self.state, Mref, self.actuation.nu, np.matrix([0., 0.]).T)
contactModel.addContact('contact_' + str(i), supportContactModel)
# Creating the cost model for a contact phase
costModel = crocoddyl.CostModelSum(self.state, self.actuation.nu)
if swingFootTask is not None:
for i in swingFootTask:
footTrack = crocoddyl.CostModelFramePlacement(self.state, i, self.actuation.nu)
costModel.addCost("footTrack_" + str(i), footTrack, 1e8)
footVel = crocoddyl.FrameMotion(i.frame, pinocchio.Motion.Zero())
impactFootVelCost = crocoddyl.CostModelFrameVelocity(self.state, footVel, self.actuation.nu)
costModel.addCost('impactVel_' + str(i.frame), impactFootVelCost, 1e6)
stateWeights = np.array([0] * 3 + [500.] * 3 + [0.01] * (self.state.nv - 6) + [10] * self.state.nv)
stateReg = crocoddyl.CostModelState(self.state,
crocoddyl.ActivationModelWeightedQuad(np.matrix(stateWeights**2).T),
self.rmodel.defaultState, self.actuation.nu)
ctrlReg = crocoddyl.CostModelControl(self.state, self.actuation.nu)
costModel.addCost("stateReg", stateReg, 1e1)
costModel.addCost("ctrlReg", ctrlReg, 1e-3)
# Creating the action model for the KKT dynamics with simpletic Euler
# integration scheme
dmodel = crocoddyl.DifferentialActionModelContactFwdDynamics(self.state, self.actuation, contactModel,
costModel)
model = crocoddyl.IntegratedActionModelEuler(dmodel, 0.)
return model
robot_model = talos_arm.model
# Create a cost model per the running and terminal action model.
state = crocoddyl.StateMultibody(robot_model)
runningCostModel = crocoddyl.CostModelSum(state)
terminalCostModel = crocoddyl.CostModelSum(state)
# Note that we need to include a cost model (i.e. set of cost functions) in
# order to fully define the action model for our optimal control problem.
# For this particular example, we formulate three running-cost functions:
# goal-tracking cost, state and control regularization; and one terminal-cost:
# goal cost. First, let's create the common cost functions.
Mref = crocoddyl.FramePlacement(
robot_model.getFrameId("gripper_left_joint"),
pinocchio.SE3(np.eye(3), np.matrix([[.0], [.0], [.4]])))
goalTrackingCost = crocoddyl.CostModelFramePlacement(state, Mref)
xRegCost = crocoddyl.CostModelState(state)
uRegCost = crocoddyl.CostModelControl(state)
# Then let's added the running and terminal cost functions
runningCostModel.addCost("gripperPose", goalTrackingCost, 1)
runningCostModel.addCost("xReg", xRegCost, 1e-4)
runningCostModel.addCost("uReg", uRegCost, 1e-4)
terminalCostModel.addCost("gripperPose", goalTrackingCost, 1)
# Next, we need to create an action model for running and terminal knots. The
# forward dynamics (computed using ABA) are implemented
# inside DifferentialActionModelFullyActuated.
actuationModel = crocoddyl.ActuationModelFull(state)
dt = 1e-3
runningModel = crocoddyl.IntegratedActionModelEuler(
stateWeightsTerm = np.array([0] * 3 + [10.] * 3 + [0.01] * (state.nv - 6) +
[100] * state.nv)
xRegCost = crocoddyl.CostModelState(
state, crocoddyl.ActivationModelWeightedQuad(stateWeights**2),
rmodel.defaultState, actuation.nu)
uRegCost = crocoddyl.CostModelControl(state, actuation.nu)
xRegTermCost = crocoddyl.CostModelState(
state, crocoddyl.ActivationModelWeightedQuad(stateWeightsTerm**2),
rmodel.defaultState, actuation.nu)
# Cost for target reaching: hand and foot
handTrackingWeights = np.array([1] * 3 + [0.0001] * 3)
Pref = crocoddyl.FramePlacement(endEffectorId,
pinocchio.SE3(np.eye(3), target))
handTrackingCost = crocoddyl.CostModelFramePlacement(
state, crocoddyl.ActivationModelWeightedQuad(handTrackingWeights**2), Pref,
actuation.nu)
footTrackingWeights = np.array([1, 1, 0.1] + [1.] * 3)
Pref = crocoddyl.FramePlacement(
leftFootId, pinocchio.SE3(np.eye(3), np.array([0., 0.4, 0.])))
footTrackingCost1 = crocoddyl.CostModelFramePlacement(
state, crocoddyl.ActivationModelWeightedQuad(footTrackingWeights**2), Pref,
actuation.nu)
Pref = crocoddyl.FramePlacement(
leftFootId, pinocchio.SE3(np.eye(3), np.array([0.3, 0.15, 0.35])))
footTrackingCost2 = crocoddyl.CostModelFramePlacement(
state, crocoddyl.ActivationModelWeightedQuad(footTrackingWeights**2), Pref,
actuation.nu)
# Cost for CoM reference
[100] * state.nv)
xRegCost = crocoddyl.CostModelState(
state, crocoddyl.ActivationModelWeightedQuad(stateWeights**2),
rmodel.defaultState, actuation.nu)
uRegCost = crocoddyl.CostModelControl(state, actuation.nu)
xRegTermCost = crocoddyl.CostModelState(
state, crocoddyl.ActivationModelWeightedQuad(stateWeightsTerm**2),
rmodel.defaultState, actuation.nu)
# Cost for target reaching
goaltrackingWeights = np.array([1] * 3 + [0.0001] * 3)
framePoseEff = pinocchio.SE3.Identity()
framePoseEff.translation = target
Pref = crocoddyl.FramePlacement(endEffectorId, framePoseEff)
goalTrackingCost = crocoddyl.CostModelFramePlacement(
state, crocoddyl.ActivationModelWeightedQuad(goaltrackingWeights**2), Pref,
actuation.nu)
# Cost for CoM reference
comTrack = crocoddyl.CostModelCoMPosition(state, comRef, actuation.nu)
# Create cost model per each action model
runningCostModel = crocoddyl.CostModelSum(state, actuation.nu)
terminalCostModel = crocoddyl.CostModelSum(state, actuation.nu)
# Then let's added the running and terminal cost functions
runningCostModel.addCost("gripperPose", goalTrackingCost, 1e2)
runningCostModel.addCost("stateReg", xRegCost, 1e-3)
runningCostModel.addCost("ctrlReg", uRegCost, 1e-4)
runningCostModel.addCost("limitCost", limitCost, 1e3)
