def get_default_randomizer_pend() -> DomainRandomizer:
"""
Get the default randomizer for the `PendulumSim`.
:return: randomizer based on the nominal domain parameter values
"""
from pyrado.environments.pysim.pendulum import PendulumSim
dp_nom = PendulumSim.get_nominal_domain_param()
return DomainRandomizer(
NormalDomainParam(name='g',
mean=dp_nom['g'],
std=dp_nom['g'] / 10,
clip_lo=1e-3),
NormalDomainParam(name='m_pole',
mean=dp_nom['m_pole'],
std=dp_nom['m_pole'] / 10,
clip_lo=1e-3),
NormalDomainParam(name='l_pole',
mean=dp_nom['l_pole'],
std=dp_nom['l_pole'] / 10,
clip_lo=1e-3),
NormalDomainParam(name='d_pole',
mean=dp_nom['d_pole'],
std=dp_nom['d_pole'] / 10,
clip_lo=1e-3),
NormalDomainParam(name='tau_max',
mean=dp_nom['tau_max'],
std=dp_nom['tau_max'] / 10,
clip_lo=1e-3))
def create_default_randomizer_pend() -> DomainRandomizer:
"""
Create the default randomizer for the `PendulumSim`.
:return: randomizer based on the nominal domain parameter values
"""
from pyrado.environments.pysim.pendulum import PendulumSim
dp_nom = PendulumSim.get_nominal_domain_param()
return DomainRandomizer(
NormalDomainParam(name="gravity_const",
mean=dp_nom["gravity_const"],
std=dp_nom["gravity_const"] / 10,
clip_lo=1e-3),
NormalDomainParam(name="pole_mass",
mean=dp_nom["pole_mass"],
std=dp_nom["pole_mass"] / 10,
clip_lo=1e-3),
NormalDomainParam(name="pole_length",
mean=dp_nom["pole_length"],
std=dp_nom["pole_length"] / 10,
clip_lo=1e-3),
NormalDomainParam(name="pole_damping",
mean=dp_nom["pole_damping"],
std=dp_nom["pole_damping"] / 10,
clip_lo=1e-3),
NormalDomainParam(name="torque_thold",
mean=dp_nom["torque_thold"],
std=dp_nom["torque_thold"] / 10,
clip_lo=1e-3),
)
# Environments
env_hparams = dict(dt=1 / 50.0, max_steps=400)
env_sim = PendulumSim(**env_hparams)
env_sim.domain_param = dict(d_pole=0)
env_sim.domain_param = dict(tau_max=4.5)
# Create a fake ground truth target domain
num_real_rollouts = 1
env_real = deepcopy(env_sim)
# Define a mapping: index - domain parameter
dp_mapping = {0: "pole_mass", 1: "pole_length"}
# Prior
dp_nom = env_sim.get_nominal_domain_param()
prior_hparam = dict(
low=to.tensor([dp_nom["pole_mass"] * 0.3, dp_nom["pole_length"] * 0.3]),
high=to.tensor([dp_nom["pole_mass"] * 1.7, dp_nom["pole_length"] * 1.7]),
)
prior = sbiutils.BoxUniform(**prior_hparam)
# Time series embedding
embedding_hparam = dict(downsampling_factor=1)
embedding = create_embedding(BayesSimEmbedding.name, env_sim.spec, **embedding_hparam)
# Behavioral policy
policy = create_pend_excitation_policy(env_sim, 1)
# Algorithm
algo_hparam = dict(