def create_home_pos_task(env_spec: EnvSpec, obs_labels: Sequence[str],
state_des: np.ndarray) -> MaskedTask:
"""
Create a task for moving the robot to safe position.
.. note::
This task was designed with an RcsPySim environment in mind, but is not restricted to these environments.
:param env_spec: environment specification
:param obs_labels: labels for selection, e.g. ['PowerGrasp_R_Y', 'PowerGrasp_R_Z']. This needs to match the
observations' names in RcsPySim
:param state_des: desired state (depends of the coordinate system). If reached, the task is over.
:return: masked task that only considers a subspace of all observations
"""
# Get the masked environment specification
spec = EnvSpec(
env_spec.obs_space,
env_spec.act_space,
env_spec.state_space.subspace(
env_spec.state_space.create_mask(obs_labels)),
)
# Create a desired state task
Q = 1e1 * np.eye(len(state_des))
R = 1e-1 * np.eye(spec.act_space.flat_dim)
rew_fcn = ExpQuadrErrRewFcn(Q, R)
task = DesStateTask(spec, state_des, rew_fcn)
# Return the masked tasks
return MaskedTask(env_spec, task, obs_labels)
def _create_task(self, task_args: dict) -> Task:
# Define the task including the reward function
state_des = task_args.get("state_des", np.array([0.0, np.pi, 0.0, 0.0]))
Q = task_args.get("Q", np.diag([3.0, 4.0, 2.0, 2.0]))
R = task_args.get("R", np.diag([5e-2]))
return RadiallySymmDesStateTask(self.spec, state_des, ExpQuadrErrRewFcn(Q, R), idcs=[1])
def _create_task(self, task_args: dict) -> Task:
# Set up task. We track the distance to the goal for both hands separately.
continuous_rew_fcn = task_args.get('continuous_rew_fcn', True)
mps_left = task_args.get('mps_left')
mps_right = task_args.get('mps_right')
if continuous_rew_fcn:
Q = np.diag([1, 1e-3])
R = 1e-4 * np.eye(self.act_space.flat_dim)
rew_fcn_factory = lambda: ExpQuadrErrRewFcn(Q, R)
else:
rew_fcn_factory = MinusOnePerStepRewFcn
succ_thold = 7.5e-2
tasks_left = [
create_goal_dist_distvel_task(self.spec, i, rew_fcn_factory(),
succ_thold)
for i in range(len(mps_left))
]
tasks_right = [
create_goal_dist_distvel_task(self.spec, i + len(mps_left),
rew_fcn_factory(), succ_thold)
for i in range(len(mps_right))
]
return ParallelTasks([
SequentialTasks(tasks_left, hold_rew_when_done=continuous_rew_fcn),
SequentialTasks(tasks_right,
hold_rew_when_done=continuous_rew_fcn),
],
hold_rew_when_done=continuous_rew_fcn)
def create_box_lift_task(env_spec: EnvSpec, continuous_rew_fcn: bool,
succ_thold: float):
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
idcs = ['Box_Z']
# Get the masked environment specification
spec = EnvSpec(
env_spec.obs_space, env_spec.act_space,
env_spec.state_space.subspace(env_spec.state_space.create_mask(idcs)))
# Create a desired state task
# state_des = np.array([0.3]) # box position is measured relative to the table
state_des = np.array([1.1]) # box position is measured world coordinates
if continuous_rew_fcn:
Q = np.diag([3e1])
R = 1e0 * np.eye(spec.act_space.flat_dim)
rew_fcn = ExpQuadrErrRewFcn(Q, R)
else:
rew_fcn = MinusOnePerStepRewFcn()
dst = DesStateTask(
spec, state_des, rew_fcn,
functools.partial(proximity_succeeded, thold_dist=succ_thold))
# Return the masked tasks
return MaskedTask(env_spec, dst, idcs)
def create_box_upper_shelve_task(env_spec: EnvSpec, continuous_rew_fcn: bool,
succ_thold: float):
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
idcs = ['Box_X', 'Box_Y', 'Box_Z', 'Box_A', 'Box_B', 'Box_C']
# Get the masked environment specification
spec = EnvSpec(
env_spec.obs_space, env_spec.act_space,
env_spec.state_space.subspace(env_spec.state_space.create_mask(idcs)))
# Create a desired state task
state_des = np.zeros(
6) # zeros since we observe the box position relative to the goal
if continuous_rew_fcn:
Q = np.diag([5e0, 5e0, 5e0, 1e-1, 1e-1, 1e-1])
R = 5e-2 * np.eye(spec.act_space.flat_dim)
rew_fcn = ExpQuadrErrRewFcn(Q, R)
else:
rew_fcn = MinusOnePerStepRewFcn
dst = DesStateTask(
spec, state_des, rew_fcn,
functools.partial(proximity_succeeded, thold_dist=succ_thold))
# Return the masked tasks
return MaskedTask(env_spec, dst, idcs)
def _create_task(self, task_args: dict) -> Task:
# Define the task including the reward function
state_des = task_args.get('state_des', np.array([0., np.pi, 0., 0.]))
Q = task_args.get('Q', np.diag([2e-1, 1., 2e-2, 5e-3]))
R = task_args.get('R', np.diag([3e-3]))
return RadiallySymmDesStateTask(self.spec,
state_des,
ExpQuadrErrRewFcn(Q, R),
idcs=[1])
def _create_task(self, task_args: dict) -> Task:
# Define the task including the reward function
state_des = task_args.get('state_des', None)
if state_des is None:
state_des = np.array([0., np.pi, 0., 0.])
Q = np.diag([3., 4., 2., 2.])
R = np.diag([5e-2])
return RadiallySymmDesStateTask(self.spec,
state_des,
ExpQuadrErrRewFcn(Q, R),
idcs=[1])
def _create_main_task(self, task_args: dict) -> Task:
# Create a DesStateTask that masks everything but the ball position
idcs = list(
range(self.state_space.flat_dim - 6,
self.state_space.flat_dim - 3)) # Cartesian ball position
spec = EnvSpec(
self.spec.obs_space, self.spec.act_space,
self.spec.state_space.subspace(
self.spec.state_space.create_mask(idcs)))
# If we do not use copy(), state_des coming from MuJoCo is a reference and updates automatically at each step.
# Note: sim.forward() + get_body_xpos() results in wrong output for state_des, as sim has not been updated to
# init_space.sample(), which is first called in reset()
if task_args.get('sparse_rew_fcn', False):
factor = task_args.get('success_bonus', 1)
# Binary final reward task
main_task = FinalRewTask(ConditionOnlyTask(
spec,
condition_fcn=self.check_ball_in_cup,
is_success_condition=True),
mode=FinalRewMode(always_positive=True),
factor=factor)
# Yield -1 on fail after the main task ist done (successfully or not)
dont_fail_after_succ_task = FinalRewTask(
GoallessTask(spec, ZeroPerStepRewFcn()),
mode=FinalRewMode(always_negative=True),
factor=factor)
# Augment the binary task with an endless dummy task, to avoid early stopping
task = SequentialTasks((main_task, dont_fail_after_succ_task))
return MaskedTask(self.spec, task, idcs)
else:
state_des = self.sim.data.get_site_xpos(
'cup_goal') # this is a reference
R_default = np.diag([
0, 0, 1, 1e-2, 1e-2, 1e-1
]) if self.num_dof == 7 else np.diag([0, 0, 1e-2, 1e-2])
rew_fcn = ExpQuadrErrRewFcn(
Q=task_args.get('Q', np.diag([
2e1, 1e-4, 2e1
])), # distance ball - cup; shouldn't move in y-direction
R=task_args.get('R',
R_default) # last joint is really unreliable
)
task = DesStateTask(spec, state_des, rew_fcn)
# Wrap the masked DesStateTask to add a bonus for the best state in the rollout
return BestStateFinalRewTask(MaskedTask(self.spec, task, idcs),
max_steps=self.max_steps,
factor=task_args.get(
'final_factor', 0.05))
def _create_task(self, task_args: dict) -> Task:
# Create a DesStateTask that masks everything but the ball position
idcs = list(
range(self.state_space.flat_dim - 3,
self.state_space.flat_dim)) # Cartesian ball position
spec = EnvSpec(
self.spec.obs_space, self.spec.act_space,
self.spec.state_space.subspace(
self.spec.state_space.create_mask(idcs)))
# If we do not use copy(), state_des coming from MuJoCo is a reference and updates automatically at each step.
# Note: sim.forward() + get_body_xpos() results in wrong output for state_des, as sim has not been updated to
# init_space.sample(), which is first called in reset()
if task_args.get('sparse_rew_fcn', False):
# Binary final reward task
task = FinalRewTask(ConditionOnlyTask(
spec,
condition_fcn=self.check_ball_in_cup,
is_success_condition=True),
mode=FinalRewMode(always_positive=True),
factor=1)
return MaskedTask(self.spec, task, idcs)
else:
# If we do not use copy(), state_des is a reference to passed body and updates automatically at each step
state_des = self.sim.data.get_site_xpos(
'cup_goal') # this is a reference
rew_fcn = ExpQuadrErrRewFcn(
Q=task_args.get('Q', np.diag([
1e1, 1e5, 2e1
])), # distance ball - cup; shouldn't move in y-direction
R=task_args.get('R',
np.diag([
1e-1, 1e-1, 1e-1, 1e-2, 1e-2, 1e-2
])) # desired joint angles and velocities
)
task = DesStateTask(spec, state_des, rew_fcn)
# Wrap the masked DesStateTask to add a bonus for the best state in the rollout
return BestStateFinalRewTask(MaskedTask(self.spec, task, idcs),
max_steps=self.max_steps,
factor=task_args.get(
'final_factor', 1.))
def create_extract_slider_task(env_spec: EnvSpec, task_args: dict,
des_state: np.ndarray):
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
idcs = ['Slider_Y']
# Get the masked environment specification
spec = EnvSpec(
env_spec.obs_space, env_spec.act_space,
env_spec.state_space.subspace(env_spec.state_space.create_mask(idcs)))
# Create a desired state task
Q = task_args.get('Q_slider', np.array([[4e1]]))
R = task_args.get('R_slider', 1e-6 * np.eye(spec.act_space.flat_dim))
rew_fcn = ExpQuadrErrRewFcn(Q, R)
dst_task = DesStateTask(spec, des_state, rew_fcn)
# Return the masked tasks
return MaskedTask(env_spec, dst_task, idcs)
def create_lifting_task(
env_spec: EnvSpec,
obs_labels: Sequence[str],
des_height: Union[float, np.ndarray],
succ_thold: float = 0.01,
) -> MaskedTask:
"""
Create a task for lifting an object.
.. note::
This task was designed with an RcsPySim environment in mind, but is not restricted to these environments.
:param env_spec: environment specification
:param obs_labels: labels for selection, e.g. ['Box_Z']. This needs to match the observations' names in RcsPySim
:param des_height: desired height of the object (depends of the coordinate system). If reached, the task is over.
:param succ_thold: once the object of interest is closer than this threshold, the task is considered successfully
:return: masked task that only considers a subspace of all observations
"""
# Get the masked environment specification
spec = EnvSpec(
env_spec.obs_space,
env_spec.act_space,
env_spec.state_space.subspace(
env_spec.state_space.create_mask(obs_labels)),
)
# Create a desired state task
state_des = np.asarray(des_height)
Q = np.diag([6e2])
R = 1e-1 * np.eye(spec.act_space.flat_dim)
rew_fcn = ExpQuadrErrRewFcn(Q, R)
task = DesStateTask(
spec, state_des, rew_fcn,
functools.partial(proximity_succeeded, thold_dist=succ_thold))
# Return the masked tasks
return MaskedTask(env_spec, task, obs_labels)
def _create_task(self, task_args: dict) -> Task:
# Define the task including the reward function
state_des = task_args.get("state_des", None)
if state_des is None:
# Get the goal position in world coordinates
p = self.get_body_position("Goal", "", "")
state_des = np.array([p[0], p[2], 0, 0, 0,
0]) # X, Z, B, Xd, Zd, Bd
# Create the individual subtasks
task_reach_goal = create_insert_task(
self.spec,
state_des,
rew_fcn=ExpQuadrErrRewFcn(
Q=np.diag([2e1, 2e1, 1e-1, 1e-2, 1e-2, 1e-2]),
R=2e-2 * np.eye(self.act_space.flat_dim)),
success_fcn=partial(proximity_succeeded,
thold_dist=0.07,
dims=[0, 1, 2]), # position and angle
)
task_ts_discrepancy = create_task_space_discrepancy_task(
self.spec,
AbsErrRewFcn(q=0.1 * np.ones(2), r=np.zeros(self.act_space.shape)))
return ParallelTasks([task_reach_goal, task_ts_discrepancy])
def _create_task(self, task_args: dict) -> Task:
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
if task_args.get('consider_velocities', False):
idcs = ['Effector_X', 'Effector_Z', 'Effector_Xd', 'Effector_Zd']
else:
idcs = ['Effector_X', 'Effector_Z']
# Get the masked environment specification
spec = EnvSpec(
self.spec.obs_space, self.spec.act_space,
self.spec.state_space.subspace(
self.spec.state_space.create_mask(idcs)))
# Get and set goal position in world coordinates for all three sub-goals
p1 = self.get_body_position('Goal1', '', '')
p2 = self.get_body_position('Goal2', '', '')
p3 = self.get_body_position('Goal3', '', '')
if task_args.get('consider_velocities', False):
Q = np.diag([1e0, 1e0, 1e-1, 1e-1])
state_des1 = np.array([p1[0], p1[2], 0, 0])
state_des2 = np.array([p2[0], p2[2], 0, 0])
state_des3 = np.array([p3[0], p3[2], 0, 0])
else:
Q = np.diag([1e0, 1e0])
state_des1 = np.array([p1[0], p1[2]])
state_des2 = np.array([p2[0], p2[2]])
state_des3 = np.array([p3[0], p3[2]])
success_fcn = functools.partial(proximity_succeeded,
thold_dist=7.5e-2,
dims=[0, 1]) # min distance = 7cm
R = np.zeros((spec.act_space.flat_dim, spec.act_space.flat_dim))
# Create the tasks
subtask_11 = FinalRewTask(DesStateTask(spec, state_des1,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_21 = FinalRewTask(DesStateTask(spec, state_des2,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_1p = ParallelTasks([subtask_11, subtask_21],
hold_rew_when_done=True,
verbose=False)
subtask_3 = FinalRewTask(DesStateTask(spec, state_des3,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_12 = FinalRewTask(DesStateTask(spec, state_des1,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_22 = FinalRewTask(DesStateTask(spec, state_des2,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_2p = ParallelTasks([subtask_12, subtask_22],
hold_rew_when_done=True,
verbose=False)
task = FinalRewTask(SequentialTasks(
[subtask_1p, subtask_3, subtask_2p],
hold_rew_when_done=True,
verbose=True),
mode=FinalRewMode(always_positive=True),
factor=2e3)
masked_task = MaskedTask(self.spec, task, idcs)
task_check_bounds = create_check_all_boundaries_task(self.spec,
penalty=1e3)
# Return the masked task and and additional task that ends the episode if the unmasked state is out of bound
return ParallelTasks([masked_task, task_check_bounds])
def _create_task(self, task_args: dict) -> Task:
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
if task_args.get("consider_velocities", False):
idcs = ["Effector_X", "Effector_Z", "Effector_Xd", "Effector_Zd"]
else:
idcs = ["Effector_X", "Effector_Z"]
# Get the masked environment specification
spec = EnvSpec(
self.spec.obs_space,
self.spec.act_space,
self.spec.state_space.subspace(
self.spec.state_space.create_mask(idcs)),
)
# Get the goal position in world coordinates for all three sub-goals
p1 = self.get_body_position("Goal1", "", "")
p2 = self.get_body_position("Goal2", "", "")
p3 = self.get_body_position("Goal3", "", "")
state_des1 = np.array([p1[0], p1[2], 0, 0])
state_des2 = np.array([p2[0], p2[2], 0, 0])
state_des3 = np.array([p3[0], p3[2], 0, 0])
if task_args.get("consider_velocities", False):
Q = np.diag([5e-1, 5e-1, 5e-3, 5e-3])
else:
Q = np.diag([1e0, 1e0])
state_des1 = state_des1[:2]
state_des2 = state_des2[:2]
state_des3 = state_des3[:2]
success_fcn = partial(proximity_succeeded,
thold_dist=7.5e-2,
dims=[0, 1]) # min distance = 7cm
R = np.zeros((spec.act_space.flat_dim, spec.act_space.flat_dim))
# Create the tasks
subtask_1 = FinalRewTask(DesStateTask(spec, state_des1,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_2 = FinalRewTask(DesStateTask(spec, state_des2,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_3 = FinalRewTask(DesStateTask(spec, state_des3,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_4 = FinalRewTask(DesStateTask(spec, state_des1,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_5 = FinalRewTask(DesStateTask(spec, state_des2,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
subtask_6 = FinalRewTask(DesStateTask(spec, state_des3,
ExpQuadrErrRewFcn(Q, R),
success_fcn),
mode=FinalRewMode(time_dependent=True))
task = FinalRewTask(
SequentialTasks(
[
subtask_1, subtask_2, subtask_3, subtask_4, subtask_5,
subtask_6
],
hold_rew_when_done=True,
verbose=True,
),
mode=FinalRewMode(always_positive=True),
factor=5e3,
)
masked_task = MaskedTask(self.spec, task, idcs)
# Additional tasks
task_check_bounds = create_check_all_boundaries_task(self.spec,
penalty=1e3)
if isinstance(self, Planar3LinkJointCtrlSim):
# Return the masked task and and additional task that ends the episode if the unmasked state is out of bound
return ParallelTasks([masked_task, task_check_bounds],
easily_satisfied=True)
else:
task_ts_discrepancy = create_task_space_discrepancy_task(
self.spec,
AbsErrRewFcn(q=0.5 * np.ones(2),
r=np.zeros(self.act_space.shape)))
# Return the masked task and and additional task that ends the episode if the unmasked state is out of bound
return ParallelTasks(
[masked_task, task_check_bounds, task_ts_discrepancy],
easily_satisfied=True)