def test_randomizer_reproducibility():
# Get the model
sdf_model = gym_ignition_models.get_model_file("ground_plane")
# Initialize the randomizers
randomizer1 = sdf.SDFRandomizer(sdf_model=sdf_model)
randomizer2 = sdf.SDFRandomizer(sdf_model=sdf_model)
randomizer3 = sdf.SDFRandomizer(sdf_model=sdf_model)
# Randomize the ground friction of all links (the ground plane collision)
frictions = randomizer1.find_xpath("*/link/collision/surface/friction")
assert len(frictions) == 1
# Get the original model string. It is parsed and then serialized without changes.
orig_model_string1 = randomizer1.sample(pretty_print=True)
orig_model_string2 = randomizer2.sample(pretty_print=True)
orig_model_string3 = randomizer3.sample(pretty_print=True)
assert orig_model_string1 == orig_model_string2 == orig_model_string3
# Do not seed #3
randomizer1.seed(42)
randomizer2.seed(42)
# Add randomizations for #1
randomizer1.new_randomization() \
.at_xpath("*/link/collision/surface/friction/ode/mu") \
.method(Method.Absolute) \
.sampled_from(Distribution.Uniform,
UniformParams(low=0, high=100)) \
.add()
# Add randomizations for #2
randomizer2.new_randomization() \
.at_xpath("*/link/collision/surface/friction/ode/mu") \
.method(Method.Absolute) \
.sampled_from(Distribution.Uniform,
UniformParams(low=0, high=100)) \
.add()
# Add randomizations for #3
randomizer3.new_randomization() \
.at_xpath("*/link/collision/surface/friction/ode/mu") \
.method(Method.Absolute) \
.sampled_from(Distribution.Uniform,
UniformParams(low=0, high=100)) \
.add()
# Process the randomizations
randomizer1.process_data()
randomizer2.process_data()
randomizer3.process_data()
for _ in range(5):
model1 = randomizer1.sample()
model2 = randomizer2.sample()
model3 = randomizer3.sample()
assert model1 == model2
assert model1 != model3
def test_randomize_missing_element():
# Get the URDF model
urdf_model = gym_ignition_models.get_model_file("pendulum")
# Convert it to a SDF string
sdf_model_string = scenario.urdffile_to_sdfstring(urdf_model)
# Write the SDF string to a temp file
sdf_model = misc.string_to_file(sdf_model_string)
# Create the randomizer
randomizer = sdf.SDFRandomizer(sdf_model=sdf_model)
# Try to randomize a missing element
with pytest.raises(RuntimeError):
# The ode/mu elements are missing
randomizer.new_randomization() \
.at_xpath("*/link/collision/surface/friction/ode/mu") \
.method(Method.Absolute) \
.sampled_from(Distribution.Uniform,
UniformParams(low=0, high=100)) \
.add()
# Add the missing friction/ode/mu element. We assume that friction exists.
frictions = randomizer.find_xpath("*/link/collision/surface/friction")
for friction in frictions:
# Create parent 'ode' first
if friction.find("ode") is None:
etree.SubElement(friction, "ode")
# Create child 'mu' after
ode = friction.find("ode")
if ode.find("mu") is None:
etree.SubElement(ode, "mu")
# Assign a dummy value to mu
mu = ode.find("mu")
mu.text = str(0)
# Apply the same randomization
randomizer.new_randomization() \
.at_xpath("*/link/collision/surface/friction/ode/mu") \
.method(Method.Absolute) \
.sampled_from(Distribution.Uniform,
UniformParams(low=0, high=100)) \
.ignore_zeros(False) \
.add()
# Process the randomization and sample a model
randomizer.process_data()
model1 = randomizer.sample(pretty_print=True)
model2 = randomizer.sample(pretty_print=True)
assert model1 != model2
def _get_sdf_randomizer(
self, task: SupportedTasks) -> randomizers.model.sdf.SDFRandomizer:
if self._sdf_randomizer is not None:
return self._sdf_randomizer
# Get the model file
urdf_model_file = cartpole.CartPole.get_model_file()
# Convert the URDF to SDF
sdf_model_string = scenario.urdffile_to_sdfstring(urdf_model_file)
# Write the SDF string to a temp file
sdf_model = utils.misc.string_to_file(sdf_model_string)
# Create and initialize the randomizer
sdf_randomizer = randomizers.model.sdf.SDFRandomizer(
sdf_model=sdf_model)
# Use the RNG of the task
sdf_randomizer.rng = task.np_random
# Randomize the mass of all links
sdf_randomizer.new_randomization().at_xpath(
"*/link/inertial/mass").method(Method.Additive).sampled_from(
Distribution.Uniform,
UniformParams(low=-0.2, high=0.2)).force_positive().add()
# Process the randomization
sdf_randomizer.process_data()
assert len(sdf_randomizer.get_active_randomizations()) > 0
# Store and return the randomizer
self._sdf_randomizer = sdf_randomizer
return self._sdf_randomizer
def test_sdf_randomizer():
# Get the URDF model
urdf_model = gym_ignition_models.get_model_file("cartpole")
# Convert it to a SDF string
sdf_model_string = scenario.urdffile_to_sdfstring(urdf_model)
# Write the SDF string to a temp file
sdf_model = misc.string_to_file(sdf_model_string)
# Create the randomizer
randomizer = sdf.SDFRandomizer(sdf_model=sdf_model)
# Get the original model string. It is parsed and then serialized without changes.
orig_model = randomizer.sample(pretty_print=True)
with pytest.raises(ValueError):
# Setting wrong distribution
randomizer.new_randomization() \
.at_xpath("*/link[@name='pole']/inertial/inertia/ixx") \
.method(Method.Additive) \
.sampled_from(Distribution.Uniform, GaussianParams(mean=0, variance=0.1)) \
.add()
# Add a uniform randomization
randomizer.new_randomization() \
.at_xpath("*/link[@name='pole']/inertial/inertia/ixx") \
.method(Method.Additive) \
.sampled_from(Distribution.Gaussian, GaussianParams(mean=0, variance=0.1)) \
.add()
randomizer.process_data()
assert len(randomizer.get_active_randomizations()) == 1
assert randomizer.sample(pretty_print=True) != orig_model
# Clean the randomizer
randomizer.clean()
assert len(randomizer.get_active_randomizations()) == 0
assert randomizer.sample(pretty_print=True) == orig_model
# Add a multi-match randomization
randomizer.new_randomization() \
.at_xpath("*/link/inertial/inertia/ixx") \
.method(Method.Coefficient) \
.sampled_from(Distribution.Uniform, UniformParams(low=0.8, high=1.2)) \
.add()
assert len(randomizer.get_active_randomizations()) == 1
# Expand the matches
randomizer.process_data()
assert len(randomizer.get_active_randomizations()) > 1
# Sample
assert randomizer.sample(pretty_print=True) != orig_model
def test_full_panda_randomization():
# Get the URDF model
urdf_model = gym_ignition_models.get_model_file("panda")
# Convert it to a SDF string
sdf_model_string = scenario.urdffile_to_sdfstring(urdf_model)
# Write the SDF string to a temp file
sdf_model = misc.string_to_file(sdf_model_string)
# Create the randomizer
randomizer = sdf.SDFRandomizer(sdf_model=sdf_model)
joint_dynamics = randomizer.find_xpath("*/joint/axis/dynamics")
assert len(joint_dynamics) > 0
# Add the friction and damping elements since they're missing in the model
for joint_dynamic in joint_dynamics:
if joint_dynamic.find("friction") is None:
etree.SubElement(joint_dynamic, "friction")
friction = joint_dynamic.find("friction")
friction.text = str(0)
if joint_dynamic.find("damping") is None:
etree.SubElement(joint_dynamic, "damping")
damping = joint_dynamic.find("damping")
damping.text = str(3)
randomization_config = {
"*/link/inertial/mass": {
# mass + U(-0.5, 0.5)
"method": Method.Additive,
"distribution": Distribution.Uniform,
"params": UniformParams(low=-0.5, high=0.5),
"ignore_zeros": True,
"force_positive": True,
},
"*/link/inertial/inertia/ixx": {
# inertia * N(1, 0.2)
"method": Method.Coefficient,
"distribution": Distribution.Gaussian,
"params": GaussianParams(mean=1.0, variance=0.2),
"ignore_zeros": True,
"force_positive": True,
},
"*/link/inertial/inertia/iyy": {
"method": Method.Coefficient,
"distribution": Distribution.Gaussian,
"params": GaussianParams(mean=1.0, variance=0.2),
"ignore_zeros": True,
"force_positive": True,
},
"*/link/inertial/inertia/izz": {
"method": Method.Coefficient,
"distribution": Distribution.Gaussian,
"params": GaussianParams(mean=1.0, variance=0.2),
"ignore_zeros": True,
"force_positive": True,
},
"*/joint/axis/dynamics/friction": {
# friction in [0, 5]
"method": Method.Absolute,
"distribution": Distribution.Uniform,
"params": UniformParams(low=0, high=5),
"ignore_zeros": False, # We initialized the value as 0
"force_positive": True,
},
"*/joint/axis/dynamics/damping": {
# damping (= 3.0) * [0.8, 1.2]
"method": Method.Coefficient,
"distribution": Distribution.Uniform,
"params": UniformParams(low=0.8, high=1.2),
"ignore_zeros": True,
"force_positive": True,
},
# TODO: */joint/axis/limit/effort
}
for xpath, config in randomization_config.items():
randomizer.new_randomization().at_xpath(xpath).method(
config["method"]).sampled_from(
config["distribution"], config["params"]).force_positive(
config["distribution"]).ignore_zeros(
config["ignore_zeros"]).add()
randomizer.process_data()
assert len(randomizer.get_active_randomizations()) > 0
randomizer.sample(pretty_print=True)