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:
# Define the task including the reward function
state_des = task_args.get("state_des", np.array([0.0, np.pi, 0.0, 0.0]))
return FinalRewTask(
RadiallySymmDesStateTask(self.spec, state_des, UnderActuatedSwingUpRewFcn(), idcs=[1]),
mode=FinalRewMode(always_negative=True),
)
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", np.array([0.0, np.pi, 0.0, 0.0]))
Q = task_args.get("Q", np.diag([5e-0, 1e1, 1e-2, 1e-2]))
R = task_args.get("R", np.diag([1e-3]))
return FinalRewTask(
RadiallySymmDesStateTask(self.spec, state_des, QuadrErrRewFcn(Q, R), idcs=[1]),
mode=FinalRewMode(state_dependent=True, time_dependent=True),
)
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_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.])
return FinalRewTask(RadiallySymmDesStateTask(
self.spec,
state_des,
UnderActuatedSwingUpRewFcn(c_act=1e-2),
idcs=[1]),
mode=FinalRewMode(always_negative=True))
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([3e-1, 5e-1, 5e-3, 1e-3]))
R = task_args.get("R", np.diag([1e-3]))
rew_fcn = QuadrErrRewFcn(Q, R)
return FinalRewTask(
RadiallySymmDesStateTask(self.spec, state_des, rew_fcn, idcs=[1]),
mode=FinalRewMode(always_negative=True),
factor=1e4,
)
def test_modulated_rew_fcn():
Q = np.eye(4)
R = np.eye(2)
s = np.array([1, 2, 3, 4])
a = np.array([0, 0])
# Modulo 2 for all selected states
idcs = [0, 1, 3]
rew_fcn = QuadrErrRewFcn(Q, R)
task = RadiallySymmDesStateTask(EnvSpec(None, None, None), np.zeros(4), rew_fcn, idcs, 2)
r = task.step_rew(s, a)
assert r == -(1**2 + 3**2)
# Different modulo factor for the selected states
idcs = [1, 3]
task = RadiallySymmDesStateTask(EnvSpec(None, None, None), np.zeros(4), rew_fcn, idcs, np.array([2, 3]))
r = task.step_rew(s, a)
assert r == -(1**2 + 3**2 + 1**2)