def cmu_humanoid_heterogeneous_forage(random_state=None):
"""Requires a CMU humanoid to find all items of a particular type in a maze."""
level = ('*******\n'
'* *\n'
'* P *\n'
'* *\n'
'* G *\n'
'* *\n'
'*******\n')
# Build a position-controlled CMU humanoid walker.
walker = cmu_humanoid.CMUHumanoidPositionControlled(
observable_options={'egocentric_camera': dict(enabled=True)})
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
maze = fixed_maze.FixedMazeWithRandomGoals(
entity_layer=level,
variations_layer=None,
num_spawns=1,
num_objects=6,
)
arena = mazes.MazeWithTargets(
maze=maze,
xy_scale=3.0,
z_height=2.0,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures,
)
task = random_goal_maze.ManyHeterogeneousGoalsMaze(
walker=walker,
maze_arena=arena,
target_builders=[
functools.partial(
target_sphere.TargetSphere,
radius=0.4,
rgb1=(0, 0.4, 0),
rgb2=(0, 0.7, 0)),
functools.partial(
target_sphere.TargetSphere,
radius=0.4,
rgb1=(0.4, 0, 0),
rgb2=(0.7, 0, 0)),
],
randomize_spawn_rotation=False,
target_type_rewards=[30., -10.],
target_type_proportions=[1, 1],
shuffle_target_builders=True,
aliveness_reward=0.01,
control_timestep=.03,
)
return composer.Environment(
time_limit=25,
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def test_termination_and_discount(self):
walker = cmu_humanoid.CMUHumanoid()
# Build a maze with rooms and targets.
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
arena = mazes.RandomMazeWithTargets(
x_cells=11,
y_cells=11,
xy_scale=3,
max_rooms=4,
room_min_size=4,
room_max_size=5,
spawns_per_room=1,
targets_per_room=3,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures,
)
task = random_goal_maze.ManyGoalsMaze(
walker=walker,
maze_arena=arena,
target_builder=functools.partial(target_sphere.TargetSphere,
radius=0.4,
rgb1=(0, 0, 0.4),
rgb2=(0, 0, 0.7)),
control_timestep=.03,
physics_timestep=.005,
)
random_state = np.random.RandomState(12345)
env = composer.Environment(task, random_state=random_state)
env.reset()
zero_action = np.zeros_like(env.physics.data.ctrl)
# Walker starts in upright position.
# Should not trigger failure termination in the first few steps.
for _ in range(5):
env.step(zero_action)
self.assertFalse(task.should_terminate_episode(env.physics))
np.testing.assert_array_equal(task.get_discount(env.physics), 1)
# Rotate the walker upside down and run the physics until it makes contact.
current_time = env.physics.data.time
walker.shift_pose(env.physics,
position=(0, 0, 10),
quaternion=(0, 1, 0, 0))
env.physics.forward()
while env.physics.data.ncon == 0:
env.physics.step()
env.physics.data.time = current_time
# Should now trigger a failure termination.
env.step(zero_action)
self.assertTrue(task.should_terminate_episode(env.physics))
def test_can_compile_mjcf(self):
# Set the wall and floor textures to match DMLab and set the skybox.
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
maze = padded_room.PaddedRoom(room_size=4, num_objects=2)
arena = mazes.MazeWithTargets(
maze=maze,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures)
mjcf.Physics.from_mjcf_model(arena.mjcf_model)
def cmu_humanoid_maze_forage(random_state=None):
"""Requires a CMU humanoid to find all items in a maze."""
# Build a position-controlled CMU humanoid walker.
walker = cmu_humanoid.CMUHumanoidPositionControlled(
observable_options={'egocentric_camera': dict(enabled=True)})
# Build a maze with rooms and targets.
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
arena = mazes.RandomMazeWithTargets(
x_cells=11,
y_cells=11,
xy_scale=3,
max_rooms=4,
room_min_size=4,
room_max_size=5,
spawns_per_room=1,
targets_per_room=3,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures,
)
# Build a task that rewards the agent for obtaining targets.
task = random_goal_maze.ManyGoalsMaze(
walker=walker,
maze_arena=arena,
target_builder=functools.partial(
target_sphere.TargetSphere,
radius=0.4,
rgb1=(0, 0, 0.4),
rgb2=(0, 0, 0.7)),
target_reward_scale=50.,
physics_timestep=0.005,
control_timestep=0.03,
)
return composer.Environment(time_limit=30,
task=task,
random_state=random_state,
strip_singleton_obs_buffer_dim=True)
def test_can_compile_mjcf(self):
# Set the wall and floor textures to match DMLab and set the skybox.
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
arena = mazes.RandomMazeWithTargets(x_cells=11,
y_cells=11,
xy_scale=3,
max_rooms=4,
room_min_size=4,
room_max_size=5,
spawns_per_room=1,
targets_per_room=3,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures)
mjcf.Physics.from_mjcf_model(arena.mjcf_model)
def test_observables(self):
walker = cmu_humanoid.CMUHumanoid()
# Build a maze with rooms and targets.
skybox_texture = labmaze_textures.SkyBox(style='sky_03')
wall_textures = labmaze_textures.WallTextures(style='style_01')
floor_textures = labmaze_textures.FloorTextures(style='style_01')
arena = mazes.RandomMazeWithTargets(
x_cells=11,
y_cells=11,
xy_scale=3,
max_rooms=4,
room_min_size=4,
room_max_size=5,
spawns_per_room=1,
targets_per_room=3,
skybox_texture=skybox_texture,
wall_textures=wall_textures,
floor_textures=floor_textures,
)
task = random_goal_maze.ManyGoalsMaze(
walker=walker,
maze_arena=arena,
target_builder=functools.partial(target_sphere.TargetSphere,
radius=0.4,
rgb1=(0, 0, 0.4),
rgb2=(0, 0, 0.7)),
control_timestep=.03,
physics_timestep=.005,
)
random_state = np.random.RandomState(12345)
env = composer.Environment(task, random_state=random_state)
timestep = env.reset()
self.assertIn('walker/joints_pos', timestep.observation)