def ant_run_gaps():
walker = ant.Ant()
# Build a corridor-shaped arena with gaps, where the sizes of the gaps and
# platforms are uniformly randomized.
arena = corr_arenas.GapsCorridor(
platform_length=distributions.Uniform(1.25, 2.5), # (0.3, 2.5)
gap_length=distributions.Uniform(0.3, 0.7), # (0.5, 1.25)
corridor_width=10,
corridor_length=250)
# Build a task that rewards the agent for running down the corridor at a
# specific velocity.
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(1.0, 0, 0),
stand_height=0.2,
contact_termination=False,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
# (Chongyi Zheng): redefine reward function
# task.get_reward = _ant_get_reward.__get__(task, task.get_reward)
return composer.Environment(time_limit=30,
task=task,
strip_singleton_obs_buffer_dim=True)
def ant_run_walls():
walker = ant.Ant()
arena = corr_arenas.WallsCorridor(wall_gap=4.,
wall_width=distributions.Uniform(1, 7),
wall_height=3.0,
corridor_width=10,
corridor_length=250,
include_initial_padding=False)
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(0.5, 0, 0),
walker_spawn_rotation=0,
stand_height=0.2,
contact_termination=False,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
# (Chongyi Zheng): redefine reward function
# task.get_reward = _ant_get_reward.__get__(task, task.get_reward)
return composer.Environment(time_limit=30,
task=task,
strip_singleton_obs_buffer_dim=True)
def ant_run(random_state=None):
walker = ant.Ant()
arena = corr_arenas.EmptyCorridor()
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(5, 0, 0),
walker_spawn_rotation=0,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
return composer.Environment(time_limit=30,
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def ant_run_long():
walker = ant.Ant()
arena = corr_arenas.EmptyCorridor(corridor_length=250,
visible_side_planes=False)
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(1, 0, 0),
walker_spawn_rotation=0,
stand_height=0.2,
contact_termination=False,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
# (Chongyi Zheng): redefine reward function
# task.get_reward = _ant_get_reward.__get__(task, task.get_reward)
return composer.Environment(time_limit=30,
task=task,
strip_singleton_obs_buffer_dim=True)
def ant_escape_bowl(random_state=None):
walker = ant.Ant()
# Build a bowl-shaped arena.
arena = bowl.Bowl(ground_size=(15., 15.),
hfield_size=(30, 30, 5),
terrain_smoothness=0.15,
terrain_bump_scale=2.0)
# Build a task that rewards the agent for being far from the origin.
task = escape.Escape(walker=walker,
arena=arena,
walker_spawn_position=(0, 0, 1.5),
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
return composer.Environment(
time_limit=30, # 20
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def ant_run_walls(random_state=None):
walker = ant.Ant()
arena = corr_arenas.WallsCorridor(wall_gap=4.,
wall_width=distributions.Uniform(1, 7),
wall_height=3.0,
corridor_width=10,
corridor_length=100,
include_initial_padding=False)
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(0.5, 0, 0),
walker_spawn_rotation=0,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
return composer.Environment(time_limit=30,
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def ant_run_gaps(random_state=None):
walker = ant.Ant()
# Build a corridor-shaped arena with gaps, where the sizes of the gaps and
# platforms are uniformly randomized.
arena = corr_arenas.GapsCorridor(
platform_length=distributions.Uniform(1.25, 2.5), # (0.3, 2.5)
gap_length=distributions.Uniform(0.3, 0.7), # (0.5, 1.25)
corridor_width=10,
corridor_length=250)
# Build a task that rewards the agent for running down the corridor at a
# specific velocity.
task = corr_tasks.RunThroughCorridor(walker=walker,
arena=arena,
walker_spawn_position=(1.0, 0, 0),
target_velocity=3.0,
physics_timestep=_PHYSICS_TIMESTEP,
control_timestep=_CONTROL_TIMESTEP)
return composer.Environment(time_limit=30,
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def test_proprioception(self):
walker = ant.Ant()
for item in walker.observables.proprioception:
self.assertIsInstance(item, observable_base.Observable)
def test_set_name(self):
name = 'fred'
walker = ant.Ant(name=name)
self.assertEqual(walker.mjcf_model.model, name)
def test_get_element_tuple_property(self, name):
attribute_value = getattr(ant.Ant(), name)
self.assertNotEmpty(attribute_value)
for item in attribute_value:
self.assertIsInstance(item, mjcf.Element)
def test_get_element_property(self, name):
attribute_value = getattr(ant.Ant(), name)
self.assertIsInstance(attribute_value, mjcf.Element)
