def create_flipping_task(
env_spec: EnvSpec,
obs_labels: Sequence[str],
des_angle_delta: float = np.pi / 2.0,
endless: bool = True,
) -> MaskedTask:
"""
Create a task for rotating 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_A']. This needs to match the observations' names in RcsPySim
:param des_angle_delta: desired angle to rotate. If reached, the task is reset, and rotating continues.
:param endless: if `True`, the task will promote endlessly repeated flipping about the desired angle, else only one
flip is desired
: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
rew_fcn = CosOfOneEleRewFcn(idx=0)
task = FlippingTask(spec, des_angle_delta, rew_fcn, endless=endless)
# Return the masked tasks
return MaskedTask(env_spec, task, obs_labels)
def create_box_flip_task(env_spec: EnvSpec, continuous_rew_fcn):
# Define the indices for selection. This needs to match the observations' names in RcsPySim.
idcs = ['Box_A']
# 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([-np.pi/2]) # box position is measured world coordinates
if continuous_rew_fcn:
# q = np.array([0./np.pi])
# r = 1e-6*np.ones(spec.act_space.flat_dim)
# rew_fcn_act = AbsErrRewFcn(q, r)
rew_fcn = CosOfOneEleRewFcn(idx=0)
# rew_fcn = CompoundRewFcn([rew_fcn_act, rew_fcn_box])
else:
rew_fcn = MinusOnePerStepRewFcn()
ef_task = EndlessFlippingTask(spec, rew_fcn, init_angle=0.)
# Return the masked tasks
return MaskedTask(env_spec, ef_task, idcs)
from pyrado.tasks.utils import proximity_succeeded
from pyrado.tasks.desired_state import DesStateTask, RadiallySymmDesStateTask
from pyrado.tasks.parallel import ParallelTasks
from pyrado.tasks.reward_functions import CompoundRewFcn, CosOfOneEleRewFcn, MinusOnePerStepRewFcn, QuadrErrRewFcn, \
ScaledExpQuadrErrRewFcn, RewFcn, PlusOnePerStepRewFcn
@pytest.fixture(scope='function')
def envspec_432():
return EnvSpec(obs_space=BoxSpace(-1, 1, 4), act_space=BoxSpace(-1, 1, 2), state_space=BoxSpace(-1, 1, 3))
@pytest.mark.parametrize(
'fcn_list, reset_args, reset_kwargs', [
([MinusOnePerStepRewFcn()], [None], [None]),
([CosOfOneEleRewFcn(0)], [None], [None]),
([QuadrErrRewFcn(np.eye(2), np.eye(1))], [None], [None]),
([MinusOnePerStepRewFcn(), QuadrErrRewFcn(Q=np.eye(2), R=np.eye(1))], [None, None], [None, None]),
], ids=['wo_args-wo_kwargs', 'w_args-wo_kwargs', 'w_args2-wo_kwargs', 'wo_args-w_kwargs'])
def test_combined_reward_function_step(fcn_list, reset_args, reset_kwargs):
# Create combined reward function
c = CompoundRewFcn(fcn_list)
# Create example state and action error
err_s, err_a = np.array([1., 2.]), np.array([3.])
# Calculate combined reward
rew = c(err_s, err_a)
assert isinstance(rew, float)
# Reset the reward functions
c.reset(reset_args, reset_kwargs)