def configure_location(self):
"""Configures correct locations for this arena"""
# Run superclass first
super().configure_location()
# Determine peg friction
friction = max(0.001, np.random.normal(self.table_friction[0], self.table_friction_std))
# Grab reference to the table body in the xml
table_subtree = self.worldbody.find(".//body[@name='{}']".format("table"))
# Define start position for drawing the line
pos = self.sample_start_pos()
# Define dirt material for markers
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.0",
"shininess": "0.0",
}
dirt = CustomMaterial(
texture="Dirt",
tex_name="dirt",
mat_name="dirt_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
# Define line(s) drawn on table
for i in range(self.num_sensors):
# If we're using two clusters, we resample the starting position and direction at the halfway point
if self.two_clusters and i == int(np.floor(self.num_sensors / 2)):
pos = self.sample_start_pos()
square_name2 = 'contact_'+str(i)
square2 = CylinderObject(
name=square_name2,
size=[self.line_width / 2, 0.001],
rgba=[1, 1, 1, 1],
density=1,
material=dirt,
friction=friction,
)
self.merge_asset(square2)
visual_c = square2.get_visual(site=True)
visual_c.set("pos", array_to_string([pos[0], pos[1], self.table_half_size[2]]))
visual_c.find("site").set("pos", [0, 0, 0.005])
visual_c.find("site").set("rgba", array_to_string([0, 0, 0, 0]))
table_subtree.append(visual_c)
sensor_name = square_name2 + "_sensor"
sensor_site_name = square_name2
self.sensor_names += [sensor_name]
self.sensor_site_names[sensor_name] = sensor_site_name
# Add to the current dirt path
pos = self.sample_path_pos(pos)
def configure_location(self):
"""Configures correct locations for this arena"""
# Run superclass first
super().configure_location()
# Define start position for drawing the line
pos = self.sample_start_pos()
# Define dirt material for markers
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.0",
"shininess": "0.0",
}
dirt = CustomMaterial(
texture="Dirt",
tex_name="dirt",
mat_name="dirt_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
shared=True,
)
# Define line(s) drawn on table
for i in range(self.num_markers):
# If we're using two clusters, we resample the starting position and direction at the halfway point
if self.two_clusters and i == int(np.floor(self.num_markers / 2)):
pos = self.sample_start_pos()
marker_name = f'contact{i}'
marker = CylinderObject(
name=marker_name,
size=[self.line_width / 2, 0.001],
rgba=[1, 1, 1, 1],
material=dirt,
obj_type="visual",
joints=None,
)
# Manually add this object to the arena xml
self.merge_assets(marker)
table = find_elements(root=self.worldbody, tags="body", attribs={"name": "table"}, return_first=True)
table.append(marker.get_obj())
# Add this marker to our saved list of all markers
self.markers.append(marker)
# Add to the current dirt path
pos = self.sample_path_pos(pos)
def __init__(
self,
cylinder_radius=(0.015, 0.03),
cylinder_length=0.13,
use_object_obs=True,
reward_shaping=True,
**kwargs
):
"""
Args:
cylinder_radius (2-tuple): low and high limits of the (uniformly sampled)
radius of the cylinder
cylinder_length (float): length of the cylinder
use_object_obs (bool): if True, include object information in the observation.
reward_shaping (bool): if True, use dense rewards
Inherits the Baxter environment; refer to other parameters described there.
"""
# initialize objects of interest
self.hole = PlateWithHoleObject()
cylinder_radius = np.random.uniform(0.015, 0.03)
self.cylinder = CylinderObject(
size_min=(cylinder_radius, cylinder_length),
size_max=(cylinder_radius, cylinder_length),
)
self.mujoco_objects = OrderedDict()
# whether to use ground-truth object states
self.use_object_obs = use_object_obs
# reward configuration
self.reward_shaping = reward_shaping
super().__init__(gripper_left=None, gripper_right=None, **kwargs)
def _load_model(self):
"""
Loads an xml model, puts it in self.model
"""
super()._load_model()
self.mujoco_robot.set_base_xpos([0, 0, 0])
# load model for table top workspace
self.mujoco_arena = TableArena(table_full_size=self.table_full_size,
table_friction=self.table_friction)
if self.use_indicator_object:
self.mujoco_arena.add_pos_indicator()
# The panda robot has a pedestal, we want to align it with the table
self.mujoco_arena.set_origin(
[0.16 + self.table_full_size[0] / 2, 0, 0])
# initialize objects of interest
# in original robosuite, a simple domain randomization is included in BoxObject implementation, and called here. We choose to discard that implementation.
cube = FullyFrictionalBoxObject(
size=self.boxobject_size,
friction=self.boxobject_friction,
density=self.boxobject_density,
rgba=[1, 0, 0, 1],
)
self.mujoco_cube = cube
goal = CylinderObject(
size=[0.03, 0.001],
rgba=[0, 1, 0, 1],
)
self.mujoco_goal = goal
self.mujoco_objects = OrderedDict([("cube", cube), ("goal", goal)])
# task includes arena, robot, and objects of interest
self.model = TableTopTask(
self.mujoco_arena,
self.mujoco_robot,
self.mujoco_objects,
initializer=self.placement_initializer,
visual_objects=['goal'],
)
self.model.place_objects()
def _load_model(self):
"""
Loads an xml model, puts it in self.model
"""
super()._load_model()
# Adjust base pose(s) accordingly
if self.env_configuration == "bimanual":
xpos = self.robots[0].robot_model.base_xpos_offset["empty"]
self.robots[0].robot_model.set_base_xpos(xpos)
else:
if self.env_configuration == "single-arm-opposed":
# Set up robots facing towards each other by rotating them from their default position
for robot, rotation in zip(self.robots,
(np.pi / 2, -np.pi / 2)):
xpos = robot.robot_model.base_xpos_offset["empty"]
rot = np.array((0, 0, rotation))
xpos = T.euler2mat(rot) @ np.array(xpos)
robot.robot_model.set_base_xpos(xpos)
robot.robot_model.set_base_ori(rot)
else: # "single-arm-parallel" configuration setting
# Set up robots parallel to each other but offset from the center
for robot, offset in zip(self.robots, (-0.25, 0.25)):
xpos = robot.robot_model.base_xpos_offset["empty"]
xpos = np.array(xpos) + np.array((0, offset, 0))
robot.robot_model.set_base_xpos(xpos)
# Add arena and robot
self.model = MujocoWorldBase()
self.mujoco_arena = EmptyArena()
if self.use_indicator_object:
self.mujoco_arena.add_pos_indicator()
self.model.merge(self.mujoco_arena)
for robot in self.robots:
self.model.merge(robot.robot_model)
# initialize objects of interest
self.hole = PlateWithHoleObject(name="hole", )
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.4",
"shininess": "0.1",
}
greenwood = CustomMaterial(
texture="WoodGreen",
tex_name="greenwood",
mat_name="greenwood_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
self.peg = CylinderObject(
name="peg",
size_min=(self.peg_radius[0], self.peg_length),
size_max=(self.peg_radius[1], self.peg_length),
material=greenwood,
rgba=[0, 1, 0, 1],
)
# Load hole object
self.hole_obj = self.hole.get_collision(site=True)
self.hole_obj.set("quat", "0 0 0.707 0.707")
self.hole_obj.set("pos", "0.11 0 0.17")
self.model.merge_asset(self.hole)
# Load peg object
self.peg_obj = self.peg.get_collision(site=True)
self.peg_obj.set("pos", array_to_string((0, 0, self.peg_length)))
self.model.merge_asset(self.peg)
# Depending on env configuration, append appropriate objects to arms
if self.env_configuration == "bimanual":
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name["left"])).append(
self.hole_obj)
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name["right"])).append(
self.peg_obj)
else:
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[1].robot_model.eef_name)).append(self.hole_obj)
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name)).append(self.peg_obj)
class TwoArmPegInHole(RobotEnv):
"""
This class corresponds to the peg-in-hole task for two robot arms.
Args:
robots (str or list of str): Specification for specific robot arm(s) to be instantiated within this env
(e.g: "Sawyer" would generate one arm; ["Panda", "Panda", "Sawyer"] would generate three robot arms)
Note: Must be either 2 single single-arm robots or 1 bimanual robot!
env_configuration (str): Specifies how to position the robots within the environment. Can be either:
:`'bimanual'`: Only applicable for bimanual robot setups. Sets up the (single) bimanual robot on the -x
side of the table
:`'single-arm-parallel'`: Only applicable for multi single arm setups. Sets up the (two) single armed
robots next to each other on the -x side of the table
:`'single-arm-opposed'`: Only applicable for multi single arm setups. Sets up the (two) single armed
robots opposed from each others on the opposite +/-y sides of the table (Default option)
controller_configs (str or list of dict): If set, contains relevant controller parameters for creating a
custom controller. Else, uses the default controller for this specific task. Should either be single
dict if same controller is to be used for all robots or else it should be a list of the same length as
"robots" param
gripper_types (str or list of str): type of gripper, used to instantiate
gripper models from gripper factory.
For this environment, setting a value other than the default (None) will raise an AssertionError, as
this environment is not meant to be used with any gripper at all.
gripper_visualizations (bool or list of bool): True if using gripper visualization.
Useful for teleoperation. Should either be single bool if gripper visualization is to be used for all
robots or else it should be a list of the same length as "robots" param
initialization_noise (dict or list of dict): Dict containing the initialization noise parameters.
The expected keys and corresponding value types are specified below:
:`'magnitude'`: The scale factor of uni-variate random noise applied to each of a robot's given initial
joint positions. Setting this value to `None` or 0.0 results in no noise being applied.
If "gaussian" type of noise is applied then this magnitude scales the standard deviation applied,
If "uniform" type of noise is applied then this magnitude sets the bounds of the sampling range
:`'type'`: Type of noise to apply. Can either specify "gaussian" or "uniform"
Should either be single dict if same noise value is to be used for all robots or else it should be a
list of the same length as "robots" param
:Note: Specifying "default" will automatically use the default noise settings.
Specifying None will automatically create the required dict with "magnitude" set to 0.0.
use_camera_obs (bool or list of bool): if True, every observation for a specific robot includes a rendered
image. Should either be single bool if camera obs value is to be used for all
robots or else it should be a list of the same length as "robots" param
use_object_obs (bool): if True, include object (cube) information in
the observation.
reward_scale (None or float): Scales the normalized reward function by the amount specified.
If None, environment reward remains unnormalized
reward_shaping (bool): if True, use dense rewards.
peg_radius (2-tuple): low and high limits of the (uniformly sampled)
radius of the peg
peg_length (float): length of the peg
use_indicator_object (bool): if True, sets up an indicator object that
is useful for debugging.
has_renderer (bool): If true, render the simulation state in
a viewer instead of headless mode.
has_offscreen_renderer (bool): True if using off-screen rendering
render_camera (str): Name of camera to render if `has_renderer` is True. Setting this value to 'None'
will result in the default angle being applied, which is useful as it can be dragged / panned by
the user using the mouse
render_collision_mesh (bool): True if rendering collision meshes in camera. False otherwise.
render_visual_mesh (bool): True if rendering visual meshes in camera. False otherwise.
control_freq (float): how many control signals to receive in every second. This sets the amount of
simulation time that passes between every action input.
horizon (int): Every episode lasts for exactly @horizon timesteps.
ignore_done (bool): True if never terminating the environment (ignore @horizon).
hard_reset (bool): If True, re-loads model, sim, and render object upon a reset call, else,
only calls sim.reset and resets all robosuite-internal variables
camera_names (str or list of str): name of camera to be rendered. Should either be single str if
same name is to be used for all cameras' rendering or else it should be a list of cameras to render.
:Note: At least one camera must be specified if @use_camera_obs is True.
:Note: To render all robots' cameras of a certain type (e.g.: "robotview" or "eye_in_hand"), use the
convention "all-{name}" (e.g.: "all-robotview") to automatically render all camera images from each
robot's camera list).
camera_heights (int or list of int): height of camera frame. Should either be single int if
same height is to be used for all cameras' frames or else it should be a list of the same length as
"camera names" param.
camera_widths (int or list of int): width of camera frame. Should either be single int if
same width is to be used for all cameras' frames or else it should be a list of the same length as
"camera names" param.
camera_depths (bool or list of bool): True if rendering RGB-D, and RGB otherwise. Should either be single
bool if same depth setting is to be used for all cameras or else it should be a list of the same length as
"camera names" param.
Raises:
AssertionError: [Gripper specified]
ValueError: [Invalid number of robots specified]
ValueError: [Invalid env configuration]
ValueError: [Invalid robots for specified env configuration]
"""
def __init__(
self,
robots,
env_configuration="single-arm-opposed",
controller_configs=None,
gripper_types=None,
gripper_visualizations=False,
initialization_noise="default",
use_camera_obs=True,
use_object_obs=True,
reward_scale=1.0,
reward_shaping=False,
peg_radius=(0.015, 0.03),
peg_length=0.13,
use_indicator_object=False,
has_renderer=False,
has_offscreen_renderer=True,
render_camera="frontview",
render_collision_mesh=False,
render_visual_mesh=True,
control_freq=10,
horizon=1000,
ignore_done=False,
hard_reset=True,
camera_names="agentview",
camera_heights=256,
camera_widths=256,
camera_depths=False,
):
# First, verify that correct number of robots are being inputted
self.env_configuration = env_configuration
self._check_robot_configuration(robots)
# Assert that the gripper type is None
assert gripper_types is None, "Tried to specify gripper other than None in TwoArmPegInHole environment!"
# reward configuration
self.reward_scale = reward_scale
self.reward_shaping = reward_shaping
# whether to use ground-truth object states
self.use_object_obs = use_object_obs
# Save peg specs
self.peg_radius = peg_radius
self.peg_length = peg_length
super().__init__(
robots=robots,
controller_configs=controller_configs,
gripper_types=gripper_types,
gripper_visualizations=gripper_visualizations,
initialization_noise=initialization_noise,
use_camera_obs=use_camera_obs,
use_indicator_object=use_indicator_object,
has_renderer=has_renderer,
has_offscreen_renderer=has_offscreen_renderer,
render_camera=render_camera,
render_collision_mesh=render_collision_mesh,
render_visual_mesh=render_visual_mesh,
control_freq=control_freq,
horizon=horizon,
ignore_done=ignore_done,
hard_reset=hard_reset,
camera_names=camera_names,
camera_heights=camera_heights,
camera_widths=camera_widths,
camera_depths=camera_depths,
)
def reward(self, action):
"""
Reward function for the task.
Sparse un-normalized reward:
- a discrete reward of 5.0 is provided if the peg is inside the plate's hole
- Note that we enforce that it's inside at an appropriate angle (cos(theta) > 0.95).
Un-normalized summed components if using reward shaping:
- Reaching: in [0, 1], to encourage the arms to approach each other
- Perpendicular Distance: in [0,1], to encourage the arms to approach each other
- Parallel Distance: in [0,1], to encourage the arms to approach each other
- Alignment: in [0, 1], to encourage having the right orientation between the peg and hole.
- Placement: in {0, 1}, nonzero if the peg is in the hole with a relatively correct alignment
Note that the final reward is normalized and scaled by reward_scale / 5.0 as
well so that the max score is equal to reward_scale
"""
reward = 0
# Right location and angle
if self._check_success():
reward = 1.0
# use a shaping reward
if self.reward_shaping:
# Grab relevant values
t, d, cos = self._compute_orientation()
# reaching reward
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
gripper_site_pos = self.sim.data.body_xpos[self.peg_body_id]
dist = np.linalg.norm(gripper_site_pos - hole_pos)
reaching_reward = 1 - np.tanh(1.0 * dist)
reward += reaching_reward
# Orientation reward
reward += 1 - np.tanh(d)
reward += 1 - np.tanh(np.abs(t))
reward += cos
# if we're not reward shaping, we need to scale our sparse reward so that the max reward is identical
# to its dense version
else:
reward *= 5.0
if self.reward_scale is not None:
reward *= self.reward_scale / 5.0
return reward
def _load_model(self):
"""
Loads an xml model, puts it in self.model
"""
super()._load_model()
# Adjust base pose(s) accordingly
if self.env_configuration == "bimanual":
xpos = self.robots[0].robot_model.base_xpos_offset["empty"]
self.robots[0].robot_model.set_base_xpos(xpos)
else:
if self.env_configuration == "single-arm-opposed":
# Set up robots facing towards each other by rotating them from their default position
for robot, rotation in zip(self.robots,
(np.pi / 2, -np.pi / 2)):
xpos = robot.robot_model.base_xpos_offset["empty"]
rot = np.array((0, 0, rotation))
xpos = T.euler2mat(rot) @ np.array(xpos)
robot.robot_model.set_base_xpos(xpos)
robot.robot_model.set_base_ori(rot)
else: # "single-arm-parallel" configuration setting
# Set up robots parallel to each other but offset from the center
for robot, offset in zip(self.robots, (-0.25, 0.25)):
xpos = robot.robot_model.base_xpos_offset["empty"]
xpos = np.array(xpos) + np.array((0, offset, 0))
robot.robot_model.set_base_xpos(xpos)
# Add arena and robot
self.model = MujocoWorldBase()
self.mujoco_arena = EmptyArena()
if self.use_indicator_object:
self.mujoco_arena.add_pos_indicator()
self.model.merge(self.mujoco_arena)
for robot in self.robots:
self.model.merge(robot.robot_model)
# initialize objects of interest
self.hole = PlateWithHoleObject(name="hole", )
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.4",
"shininess": "0.1",
}
greenwood = CustomMaterial(
texture="WoodGreen",
tex_name="greenwood",
mat_name="greenwood_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
self.peg = CylinderObject(
name="peg",
size_min=(self.peg_radius[0], self.peg_length),
size_max=(self.peg_radius[1], self.peg_length),
material=greenwood,
rgba=[0, 1, 0, 1],
)
# Load hole object
self.hole_obj = self.hole.get_collision(site=True)
self.hole_obj.set("quat", "0 0 0.707 0.707")
self.hole_obj.set("pos", "0.11 0 0.17")
self.model.merge_asset(self.hole)
# Load peg object
self.peg_obj = self.peg.get_collision(site=True)
self.peg_obj.set("pos", array_to_string((0, 0, self.peg_length)))
self.model.merge_asset(self.peg)
# Depending on env configuration, append appropriate objects to arms
if self.env_configuration == "bimanual":
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name["left"])).append(
self.hole_obj)
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name["right"])).append(
self.peg_obj)
else:
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[1].robot_model.eef_name)).append(self.hole_obj)
self.model.worldbody.find(".//body[@name='{}']".format(
self.robots[0].robot_model.eef_name)).append(self.peg_obj)
def _get_reference(self):
"""
Sets up references to important components. A reference is typically an
index or a list of indices that point to the corresponding elements
in a flatten array, which is how MuJoCo stores physical simulation data.
"""
super()._get_reference()
# Additional object references from this env
self.hole_body_id = self.sim.model.body_name2id("hole")
self.peg_body_id = self.sim.model.body_name2id("peg")
def _reset_internal(self):
"""
Resets simulation internal configurations.
"""
super()._reset_internal()
def _get_observation(self):
"""
Returns an OrderedDict containing observations [(name_string, np.array), ...].
Important keys:
`'robot-state'`: contains robot-centric information.
`'object-state'`: requires @self.use_object_obs to be True. Contains object-centric information.
`'image'`: requires @self.use_camera_obs to be True. Contains a rendered frame from the simulation.
`'depth'`: requires @self.use_camera_obs and @self.camera_depth to be True.
Contains a rendered depth map from the simulation
Returns:
OrderedDict: Observations from the environment
"""
di = super()._get_observation()
# low-level object information
if self.use_object_obs:
# Get robot prefix
if self.env_configuration == "bimanual":
pr0 = self.robots[0].robot_model.naming_prefix + "left_"
pr1 = self.robots[0].robot_model.naming_prefix + "right_"
else:
pr0 = self.robots[0].robot_model.naming_prefix
pr1 = self.robots[1].robot_model.naming_prefix
# position and rotation of peg and hole
hole_pos = np.array(self.sim.data.body_xpos[self.hole_body_id])
hole_quat = T.convert_quat(
self.sim.data.body_xquat[self.hole_body_id], to="xyzw")
di["hole_pos"] = hole_pos
di["hole_quat"] = hole_quat
peg_pos = np.array(self.sim.data.body_xpos[self.peg_body_id])
peg_quat = T.convert_quat(
self.sim.data.body_xquat[self.peg_body_id], to="xyzw")
di["peg_to_hole"] = peg_pos - hole_pos
di["peg_quat"] = peg_quat
# Relative orientation parameters
t, d, cos = self._compute_orientation()
di["angle"] = cos
di["t"] = t
di["d"] = d
di["object-state"] = np.concatenate([
di["hole_pos"],
di["hole_quat"],
di["peg_to_hole"],
di["peg_quat"],
[di["angle"]],
[di["t"]],
[di["d"]],
])
return di
def _check_success(self):
"""
Check if peg is successfully aligned and placed within the hole
Returns:
bool: True if peg is placed in hole correctly
"""
t, d, cos = self._compute_orientation()
return d < 0.06 and -0.12 <= t <= 0.14 and cos > 0.95
def _compute_orientation(self):
"""
Helper function to return the relative positions between the hole and the peg.
In particular, the intersection of the line defined by the peg and the plane
defined by the hole is computed; the parallel distance, perpendicular distance,
and angle are returned.
Returns:
3-tuple:
- (float): parallel distance
- (float): perpendicular distance
- (float): angle
"""
peg_mat = self.sim.data.body_xmat[self.peg_body_id]
peg_mat.shape = (3, 3)
peg_pos = self.sim.data.body_xpos[self.peg_body_id]
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
hole_mat = self.sim.data.body_xmat[self.hole_body_id]
hole_mat.shape = (3, 3)
v = peg_mat @ np.array([0, 0, 1])
v = v / np.linalg.norm(v)
center = hole_pos + hole_mat @ np.array([0.1, 0, 0])
t = (center - peg_pos) @ v / (np.linalg.norm(v)**2)
d = np.linalg.norm(np.cross(v, peg_pos - center)) / np.linalg.norm(v)
hole_normal = hole_mat @ np.array([0, 0, 1])
return (
t,
d,
abs(
np.dot(hole_normal, v) / np.linalg.norm(hole_normal) /
np.linalg.norm(v)),
)
def _peg_pose_in_hole_frame(self):
"""
A helper function that takes in a named data field and returns the pose of that
object in the base frame.
Returns:
np.array: (4,4) matrix corresponding to the pose of the peg in the hole frame
"""
# World frame
peg_pos_in_world = self.sim.data.get_body_xpos("peg")
peg_rot_in_world = self.sim.data.get_body_xmat("peg").reshape((3, 3))
peg_pose_in_world = T.make_pose(peg_pos_in_world, peg_rot_in_world)
# World frame
hole_pos_in_world = self.sim.data.get_body_xpos("hole")
hole_rot_in_world = self.sim.data.get_body_xmat("hole").reshape((3, 3))
hole_pose_in_world = T.make_pose(hole_pos_in_world, hole_rot_in_world)
world_pose_in_hole = T.pose_inv(hole_pose_in_world)
peg_pose_in_hole = T.pose_in_A_to_pose_in_B(peg_pose_in_world,
world_pose_in_hole)
return peg_pose_in_hole
def _check_robot_configuration(self, robots):
"""
Sanity check to make sure the inputted robots and configuration is acceptable
Args:
robots (str or list of str): Robots to instantiate within this env
"""
robots = robots if type(robots) == list or type(robots) == tuple else [
robots
]
if self.env_configuration == "single-arm-opposed" or self.env_configuration == "single-arm-parallel":
# Specifically two robots should be inputted!
is_bimanual = False
if type(robots) is not list or len(robots) != 2:
raise ValueError(
"Error: Exactly two single-armed robots should be inputted "
"for this task configuration!")
elif self.env_configuration == "bimanual":
is_bimanual = True
# Specifically one robot should be inputted!
if type(robots) is list and len(robots) != 1:
raise ValueError(
"Error: Exactly one bimanual robot should be inputted "
"for this task configuration!")
else:
# This is an unknown env configuration, print error
raise ValueError(
"Error: Unknown environment configuration received. Only 'bimanual',"
"'single-arm-parallel', and 'single-arm-opposed' are supported. Got: {}"
.format(self.env_configuration))
# Lastly, check to make sure all inputted robot names are of their correct type (bimanual / not bimanual)
for robot in robots:
if check_bimanual(robot) != is_bimanual:
raise ValueError(
"Error: For {} configuration, expected bimanual check to return {}; "
"instead, got {}.".format(self.env_configuration,
is_bimanual,
check_bimanual(robot)))
def __init__(
self,
name,
box1_size=(0.025, 0.025, 0.025),
box2_size=(0.025, 0.025, 0.0125),
use_texture=True,
):
# Set box sizes
self.box1_size = np.array(box1_size)
self.box2_size = np.array(box2_size)
# Set texture attributes
self.use_texture = use_texture
self.box1_material = None
self.box2_material = None
self.box1_rgba = RED
self.box2_rgba = BLUE
# Define materials we want to use for this object
if self.use_texture:
# Remove RGBAs
self.box1_rgba = None
self.box2_rgba = None
# Set materials for each box
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "3 3",
"specular": "0.4",
"shininess": "0.1",
}
self.box1_material = CustomMaterial(
texture="WoodRed",
tex_name="box1_tex",
mat_name="box1_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
self.box2_material = CustomMaterial(
texture="WoodBlue",
tex_name="box2_tex",
mat_name="box2_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
# Create objects
objects = []
for i, (size, mat, rgba) in enumerate(
zip(
(self.box1_size, self.box2_size),
(self.box1_material, self.box2_material),
(self.box1_rgba, self.box2_rgba),
)):
objects.append(
BoxObject(
name=f"box{i + 1}",
size=size,
rgba=rgba,
material=mat,
))
# Also add hinge for visualization
objects.append(
CylinderObject(
name="hinge",
size=np.array([
min(self.box1_size[2], self.box2_size[2]) / 5.0,
min(self.box1_size[0], self.box2_size[0])
]),
rgba=[0.5, 0.5, 0, 1],
obj_type="visual",
))
# Define hinge joint
rel_hinge_pos = [self.box2_size[0], 0, -self.box2_size[2]
] # want offset in all except y-axis
hinge_joint = {
"name": "box_hinge",
"type": "hinge",
"axis": "0 1 0", # y-axis hinge
"pos": array_to_string(rel_hinge_pos),
"stiffness": "0.0001",
"limited": "true",
"range": "0 1.57",
}
# Define positions -- second box should lie on top of first box with edge aligned at hinge joint
# Hinge visualizer should be aligned at hinge joint location
positions = [
np.zeros(3), # First box is centered at top-level body anyways
np.array([
-(self.box2_size[0] - self.box1_size[0]), 0,
self.box1_size[2] + self.box2_size[2]
]),
np.array(rel_hinge_pos),
]
quats = [
None, # Default quaternion for box 1
None, # Default quaternion for box 2
[0.707, 0.707, 0, 0], # Rotated 90 deg about x-axis
]
# Define parents -- which body each is aligned to
parents = [
None, # box 1 attached to top-level body
objects[0].root_body, # box 2 attached to box 1
objects[1].root_body, # hinge attached to box 2
]
# Run super init
super().__init__(
name=name,
objects=objects,
object_locations=positions,
object_quats=quats,
object_parents=parents,
body_joints={objects[1].root_body: [hinge_joint]},
)
def _load_model(self):
"""
Loads an xml model, puts it in self.model
"""
super()._load_model()
# Adjust base pose(s) accordingly
if self.env_configuration == "bimanual":
xpos = self.robots[0].robot_model.base_xpos_offset["empty"]
self.robots[0].robot_model.set_base_xpos(xpos)
else:
if self.env_configuration == "single-arm-opposed":
# Set up robots facing towards each other by rotating them from their default position
for robot, rotation in zip(self.robots,
(np.pi / 2, -np.pi / 2)):
xpos = robot.robot_model.base_xpos_offset["empty"]
rot = np.array((0, 0, rotation))
xpos = T.euler2mat(rot) @ np.array(xpos)
robot.robot_model.set_base_xpos(xpos)
robot.robot_model.set_base_ori(rot)
else: # "single-arm-parallel" configuration setting
# Set up robots parallel to each other but offset from the center
for robot, offset in zip(self.robots, (-0.25, 0.25)):
xpos = robot.robot_model.base_xpos_offset["empty"]
xpos = np.array(xpos) + np.array((0, offset, 0))
robot.robot_model.set_base_xpos(xpos)
# Add arena and robot
mujoco_arena = EmptyArena()
# Arena always gets set to zero origin
mujoco_arena.set_origin([0, 0, 0])
# Modify default agentview camera
mujoco_arena.set_camera(camera_name="agentview",
pos=[
1.0666432116509934, 1.4903257668114777e-08,
2.0563394967349096
],
quat=[
0.6530979871749878, 0.27104058861732483,
0.27104055881500244, 0.6530978679656982
])
# initialize objects of interest
self.hole = PlateWithHoleObject(name="hole")
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.4",
"shininess": "0.1",
}
greenwood = CustomMaterial(
texture="WoodGreen",
tex_name="greenwood",
mat_name="greenwood_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
self.peg = CylinderObject(
name="peg",
size_min=(self.peg_radius[0], self.peg_length),
size_max=(self.peg_radius[1], self.peg_length),
material=greenwood,
rgba=[0, 1, 0, 1],
joints=None,
)
# Load hole object
hole_obj = self.hole.get_obj()
hole_obj.set("quat", "0 0 0.707 0.707")
hole_obj.set("pos", "0.11 0 0.17")
# Load peg object
peg_obj = self.peg.get_obj()
peg_obj.set("pos", array_to_string((0, 0, self.peg_length)))
# Append appropriate objects to arms
if self.env_configuration == "bimanual":
r_eef, l_eef = [
self.robots[0].robot_model.eef_name[arm]
for arm in self.robots[0].arms
]
r_model, l_model = [
self.robots[0].robot_model, self.robots[0].robot_model
]
else:
r_eef, l_eef = [
robot.robot_model.eef_name for robot in self.robots
]
r_model, l_model = [
self.robots[0].robot_model, self.robots[1].robot_model
]
r_body = find_elements(root=r_model.worldbody,
tags="body",
attribs={"name": r_eef},
return_first=True)
l_body = find_elements(root=l_model.worldbody,
tags="body",
attribs={"name": l_eef},
return_first=True)
r_body.append(peg_obj)
l_body.append(hole_obj)
# task includes arena, robot, and objects of interest
# We don't add peg and hole directly since they were already appended to the robots
self.model = ManipulationTask(
mujoco_arena=mujoco_arena,
mujoco_robots=[robot.robot_model for robot in self.robots],
)
# Make sure to add relevant assets from peg and hole objects
self.model.merge_assets(self.hole)
self.model.merge_assets(self.peg)
class BaxterPegInHole(BaxterEnv):
"""
This class corresponds to the peg in hole task for the Baxter robot. There's
a cylinder attached to one gripper and a hole attached to the other one.
"""
def __init__(self,
cylinder_radius=(0.015, 0.03),
cylinder_length=0.13,
use_object_obs=True,
reward_shaping=True,
**kwargs):
"""
Args:
cylinder_radius (2-tuple): low and high limits of the (uniformly sampled)
radius of the cylinder
cylinder_length (float): length of the cylinder
use_object_obs (bool): if True, include object information in the observation.
reward_shaping (bool): if True, use dense rewards
Inherits the Baxter environment; refer to other parameters described there.
"""
# initialize objects of interest
self.hole = PlateWithHoleObject()
cylinder_radius = np.random.uniform(0.015, 0.03)
self.cylinder = CylinderObject(
size_min=(cylinder_radius, cylinder_length),
size_max=(cylinder_radius, cylinder_length),
)
self.mujoco_objects = OrderedDict()
# whether to use ground-truth object states
self.use_object_obs = use_object_obs
# reward configuration
self.reward_shaping = reward_shaping
super().__init__(gripper_left=None, gripper_right=None, **kwargs)
def _load_model(self):
"""
Loads the peg and the hole models.
"""
super()._load_model()
self.mujoco_robot.set_base_xpos([0, 0, 0])
# Add arena and robot
self.model = MujocoWorldBase()
self.arena = EmptyArena()
if self.use_indicator_object:
self.arena.add_pos_indicator()
self.model.merge(self.arena)
self.model.merge(self.mujoco_robot)
# Load hole object
self.hole_obj = self.hole.get_collision(name="hole", site=True)
self.hole_obj.set("quat", "0 0 0.707 0.707")
self.hole_obj.set("pos", "0.11 0 0.18")
self.model.merge_asset(self.hole)
self.model.worldbody.find(".//body[@name='left_hand']").append(
self.hole_obj)
# Load cylinder object
self.cyl_obj = self.cylinder.get_collision(name="cylinder", site=True)
self.cyl_obj.set("pos", "0 0 0.15")
self.model.merge_asset(self.cylinder)
self.model.worldbody.find(".//body[@name='right_hand']").append(
self.cyl_obj)
self.model.worldbody.find(".//geom[@name='cylinder']").set(
"rgba", "0 1 0 1")
def _get_reference(self):
"""
Sets up references to important components. A reference is typically an
index or a list of indices that point to the corresponding elements
in a flattened array, which is how MuJoCo stores physical simulation data.
"""
super()._get_reference()
self.hole_body_id = self.sim.model.body_name2id("hole")
self.cyl_body_id = self.sim.model.body_name2id("cylinder")
def _reset_internal(self):
"""
Resets simulation internal configurations.
"""
super()._reset_internal()
def _compute_orientation(self):
"""
Helper function to return the relative positions between the hole and the peg.
In particular, the intersection of the line defined by the peg and the plane
defined by the hole is computed; the parallel distance, perpendicular distance,
and angle are returned.
"""
cyl_mat = self.sim.data.body_xmat[self.cyl_body_id]
cyl_mat.shape = (3, 3)
cyl_pos = self.sim.data.body_xpos[self.cyl_body_id]
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
hole_mat = self.sim.data.body_xmat[self.hole_body_id]
hole_mat.shape = (3, 3)
v = cyl_mat @ np.array([0, 0, 1])
v = v / np.linalg.norm(v)
center = hole_pos + hole_mat @ np.array([0.1, 0, 0])
t = (center - cyl_pos) @ v / (np.linalg.norm(v)**2)
d = np.linalg.norm(np.cross(v, cyl_pos - center)) / np.linalg.norm(v)
hole_normal = hole_mat @ np.array([0, 0, 1])
return (
t,
d,
abs(
np.dot(hole_normal, v) / np.linalg.norm(hole_normal) /
np.linalg.norm(v)),
)
def reward(self, action):
"""
Reward function for the task.
The sparse reward is 0 if the peg is outside the hole, and 1 if it's inside.
We enforce that it's inside at an appropriate angle (cos(theta) > 0.95).
The dense reward has four components.
Reaching: in [0, 1], to encourage the arms to get together.
Perpendicular and parallel distance: in [0,1], for the same purpose.
Cosine of the angle: in [0, 1], to encourage having the right orientation.
"""
reward = 0
t, d, cos = self._compute_orientation()
# Right location and angle
if d < 0.06 and t >= -0.12 and t <= 0.14 and cos > 0.95:
reward = 1
# use a shaping reward
if self.reward_shaping:
# reaching reward
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
gripper_site_pos = self.sim.data.body_xpos[self.cyl_body_id]
dist = np.linalg.norm(gripper_site_pos - hole_pos)
reaching_reward = 1 - np.tanh(1.0 * dist)
reward += reaching_reward
# Orientation reward
reward += 1 - np.tanh(d)
reward += 1 - np.tanh(np.abs(t))
reward += cos
return reward
def _peg_pose_in_hole_frame(self):
"""
A helper function that takes in a named data field and returns the pose of that
object in the base frame.
"""
# World frame
peg_pos_in_world = self.sim.data.get_body_xpos("cylinder")
peg_rot_in_world = self.sim.data.get_body_xmat("cylinder").reshape(
(3, 3))
peg_pose_in_world = T.make_pose(peg_pos_in_world, peg_rot_in_world)
# World frame
hole_pos_in_world = self.sim.data.get_body_xpos("hole")
hole_rot_in_world = self.sim.data.get_body_xmat("hole").reshape((3, 3))
hole_pose_in_world = T.make_pose(hole_pos_in_world, hole_rot_in_world)
world_pose_in_hole = T.pose_inv(hole_pose_in_world)
peg_pose_in_hole = T.pose_in_A_to_pose_in_B(peg_pose_in_world,
world_pose_in_hole)
return peg_pose_in_hole
def _get_observation(self):
"""
Returns an OrderedDict containing observations [(name_string, np.array), ...].
Important keys:
robot-state: contains robot-centric information.
object-state: requires @self.use_object_obs to be True.
contains object-centric information.
image: requires @self.use_camera_obs to be True.
contains a rendered frame from the simulation.
depth: requires @self.use_camera_obs and @self.camera_depth to be True.
contains a rendered depth map from the simulation
"""
di = super()._get_observation()
# camera observations
if self.use_camera_obs:
camera_obs = self.sim.render(
camera_name=self.camera_name,
width=self.camera_width,
height=self.camera_height,
depth=self.camera_depth,
)
if self.camera_depth:
di["image"], di["depth"] = camera_obs
else:
di["image"] = camera_obs
# low-level object information
if self.use_object_obs:
# position and rotation of cylinder and hole
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
hole_quat = T.convert_quat(
self.sim.data.body_xquat[self.hole_body_id], to="xyzw")
di["hole_pos"] = hole_pos
di["hole_quat"] = hole_quat
cyl_pos = self.sim.data.body_xpos[self.cyl_body_id]
cyl_quat = T.convert_quat(
self.sim.data.body_xquat[self.cyl_body_id], to="xyzw")
di["cyl_to_hole"] = cyl_pos - hole_pos
di["cyl_quat"] = cyl_quat
# Relative orientation parameters
t, d, cos = self._compute_orientation()
di["angle"] = cos
di["t"] = t
di["d"] = d
di["object-state"] = np.concatenate([
di["hole_pos"],
di["hole_quat"],
di["cyl_to_hole"],
di["cyl_quat"],
[di["angle"]],
[di["t"]],
[di["d"]],
])
return di
def _check_contact(self):
"""
Returns True if gripper is in contact with an object.
"""
collision = False
contact_geoms = (self.gripper_right.contact_geoms() +
self.gripper_left.contact_geoms())
for contact in self.sim.data.contact[:self.sim.data.ncon]:
if (self.sim.model.geom_id2name(contact.geom1) in contact_geoms
or self.sim.model.geom_id2name(
contact.geom2) in contact_geoms):
collision = True
break
return collision
def _check_success(self):
"""
Returns True if task is successfully completed.
"""
t, d, cos = self._compute_orientation()
return d < 0.06 and t >= -0.12 and t <= 0.14 and cos > 0.95
class TwoArmPegInHole(TwoArmEnv):
"""
This class corresponds to the peg-in-hole task for two robot arms.
Args:
robots (str or list of str): Specification for specific robot arm(s) to be instantiated within this env
(e.g: "Sawyer" would generate one arm; ["Panda", "Panda", "Sawyer"] would generate three robot arms)
Note: Must be either 2 single single-arm robots or 1 bimanual robot!
env_configuration (str): Specifies how to position the robots within the environment. Can be either:
:`'bimanual'`: Only applicable for bimanual robot setups. Sets up the (single) bimanual robot on the -x
side of the table
:`'single-arm-parallel'`: Only applicable for multi single arm setups. Sets up the (two) single armed
robots next to each other on the -x side of the table
:`'single-arm-opposed'`: Only applicable for multi single arm setups. Sets up the (two) single armed
robots opposed from each others on the opposite +/-y sides of the table.
Note that "default" corresponds to either "bimanual" if a bimanual robot is used or "single-arm-opposed" if two
single-arm robots are used.
controller_configs (str or list of dict): If set, contains relevant controller parameters for creating a
custom controller. Else, uses the default controller for this specific task. Should either be single
dict if same controller is to be used for all robots or else it should be a list of the same length as
"robots" param
gripper_types (str or list of str): type of gripper, used to instantiate gripper models from gripper factory.
For this environment, setting a value other than the default (None) will raise an AssertionError, as
this environment is not meant to be used with any gripper at all.
initialization_noise (dict or list of dict): Dict containing the initialization noise parameters.
The expected keys and corresponding value types are specified below:
:`'magnitude'`: The scale factor of uni-variate random noise applied to each of a robot's given initial
joint positions. Setting this value to `None` or 0.0 results in no noise being applied.
If "gaussian" type of noise is applied then this magnitude scales the standard deviation applied,
If "uniform" type of noise is applied then this magnitude sets the bounds of the sampling range
:`'type'`: Type of noise to apply. Can either specify "gaussian" or "uniform"
Should either be single dict if same noise value is to be used for all robots or else it should be a
list of the same length as "robots" param
:Note: Specifying "default" will automatically use the default noise settings.
Specifying None will automatically create the required dict with "magnitude" set to 0.0.
use_camera_obs (bool or list of bool): if True, every observation for a specific robot includes a rendered
image. Should either be single bool if camera obs value is to be used for all
robots or else it should be a list of the same length as "robots" param
use_object_obs (bool): if True, include object (cube) information in
the observation.
reward_scale (None or float): Scales the normalized reward function by the amount specified.
If None, environment reward remains unnormalized
reward_shaping (bool): if True, use dense rewards.
peg_radius (2-tuple): low and high limits of the (uniformly sampled)
radius of the peg
peg_length (float): length of the peg
has_renderer (bool): If true, render the simulation state in
a viewer instead of headless mode.
has_offscreen_renderer (bool): True if using off-screen rendering
render_camera (str): Name of camera to render if `has_renderer` is True. Setting this value to 'None'
will result in the default angle being applied, which is useful as it can be dragged / panned by
the user using the mouse
render_collision_mesh (bool): True if rendering collision meshes in camera. False otherwise.
render_visual_mesh (bool): True if rendering visual meshes in camera. False otherwise.
render_gpu_device_id (int): corresponds to the GPU device id to use for offscreen rendering.
Defaults to -1, in which case the device will be inferred from environment variables
(GPUS or CUDA_VISIBLE_DEVICES).
control_freq (float): how many control signals to receive in every second. This sets the amount of
simulation time that passes between every action input.
horizon (int): Every episode lasts for exactly @horizon timesteps.
ignore_done (bool): True if never terminating the environment (ignore @horizon).
hard_reset (bool): If True, re-loads model, sim, and render object upon a reset call, else,
only calls sim.reset and resets all robosuite-internal variables
camera_names (str or list of str): name of camera to be rendered. Should either be single str if
same name is to be used for all cameras' rendering or else it should be a list of cameras to render.
:Note: At least one camera must be specified if @use_camera_obs is True.
:Note: To render all robots' cameras of a certain type (e.g.: "robotview" or "eye_in_hand"), use the
convention "all-{name}" (e.g.: "all-robotview") to automatically render all camera images from each
robot's camera list).
camera_heights (int or list of int): height of camera frame. Should either be single int if
same height is to be used for all cameras' frames or else it should be a list of the same length as
"camera names" param.
camera_widths (int or list of int): width of camera frame. Should either be single int if
same width is to be used for all cameras' frames or else it should be a list of the same length as
"camera names" param.
camera_depths (bool or list of bool): True if rendering RGB-D, and RGB otherwise. Should either be single
bool if same depth setting is to be used for all cameras or else it should be a list of the same length as
"camera names" param.
camera_segmentations (None or str or list of str or list of list of str): Camera segmentation(s) to use
for each camera. Valid options are:
`None`: no segmentation sensor used
`'instance'`: segmentation at the class-instance level
`'class'`: segmentation at the class level
`'element'`: segmentation at the per-geom level
If not None, multiple types of segmentations can be specified. A [list of str / str or None] specifies
[multiple / a single] segmentation(s) to use for all cameras. A list of list of str specifies per-camera
segmentation setting(s) to use.
Raises:
AssertionError: [Gripper specified]
ValueError: [Invalid number of robots specified]
ValueError: [Invalid env configuration]
ValueError: [Invalid robots for specified env configuration]
"""
def __init__(
self,
robots,
env_configuration="default",
controller_configs=None,
gripper_types=None,
initialization_noise="default",
use_camera_obs=True,
use_object_obs=True,
reward_scale=1.0,
reward_shaping=False,
peg_radius=(0.015, 0.03),
peg_length=0.13,
has_renderer=False,
has_offscreen_renderer=True,
render_camera="frontview",
render_collision_mesh=False,
render_visual_mesh=True,
render_gpu_device_id=-1,
control_freq=20,
horizon=1000,
ignore_done=False,
hard_reset=True,
camera_names="agentview",
camera_heights=256,
camera_widths=256,
camera_depths=False,
camera_segmentations=None, # {None, instance, class, element}
renderer="mujoco",
renderer_config=None,
):
# Assert that the gripper type is None
assert gripper_types is None, "Tried to specify gripper other than None in TwoArmPegInHole environment!"
# reward configuration
self.reward_scale = reward_scale
self.reward_shaping = reward_shaping
# whether to use ground-truth object states
self.use_object_obs = use_object_obs
# Save peg specs
self.peg_radius = peg_radius
self.peg_length = peg_length
super().__init__(
robots=robots,
env_configuration=env_configuration,
controller_configs=controller_configs,
mount_types="default",
gripper_types=gripper_types,
initialization_noise=initialization_noise,
use_camera_obs=use_camera_obs,
has_renderer=has_renderer,
has_offscreen_renderer=has_offscreen_renderer,
render_camera=render_camera,
render_collision_mesh=render_collision_mesh,
render_visual_mesh=render_visual_mesh,
render_gpu_device_id=render_gpu_device_id,
control_freq=control_freq,
horizon=horizon,
ignore_done=ignore_done,
hard_reset=hard_reset,
camera_names=camera_names,
camera_heights=camera_heights,
camera_widths=camera_widths,
camera_depths=camera_depths,
camera_segmentations=camera_segmentations,
renderer=renderer,
renderer_config=renderer_config,
)
def reward(self, action=None):
"""
Reward function for the task.
Sparse un-normalized reward:
- a discrete reward of 5.0 is provided if the peg is inside the plate's hole
- Note that we enforce that it's inside at an appropriate angle (cos(theta) > 0.95).
Un-normalized summed components if using reward shaping:
- Reaching: in [0, 1], to encourage the arms to approach each other
- Perpendicular Distance: in [0,1], to encourage the arms to approach each other
- Parallel Distance: in [0,1], to encourage the arms to approach each other
- Alignment: in [0, 1], to encourage having the right orientation between the peg and hole.
- Placement: in {0, 1}, nonzero if the peg is in the hole with a relatively correct alignment
Note that the final reward is normalized and scaled by reward_scale / 5.0 as
well so that the max score is equal to reward_scale
"""
reward = 0
# Right location and angle
if self._check_success():
reward = 1.0
# use a shaping reward
if self.reward_shaping:
# Grab relevant values
t, d, cos = self._compute_orientation()
# reaching reward
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
gripper_site_pos = self.sim.data.body_xpos[self.peg_body_id]
dist = np.linalg.norm(gripper_site_pos - hole_pos)
reaching_reward = 1 - np.tanh(1.0 * dist)
reward += reaching_reward
# Orientation reward
reward += 1 - np.tanh(d)
reward += 1 - np.tanh(np.abs(t))
reward += cos
# if we're not reward shaping, scale sparse reward so that the max reward is identical to its dense version
else:
reward *= 5.0
if self.reward_scale is not None:
reward *= self.reward_scale / 5.0
return reward
def _load_model(self):
"""
Loads an xml model, puts it in self.model
"""
super()._load_model()
# Adjust base pose(s) accordingly
if self.env_configuration == "bimanual":
xpos = self.robots[0].robot_model.base_xpos_offset["empty"]
self.robots[0].robot_model.set_base_xpos(xpos)
else:
if self.env_configuration == "single-arm-opposed":
# Set up robots facing towards each other by rotating them from their default position
for robot, rotation in zip(self.robots,
(np.pi / 2, -np.pi / 2)):
xpos = robot.robot_model.base_xpos_offset["empty"]
rot = np.array((0, 0, rotation))
xpos = T.euler2mat(rot) @ np.array(xpos)
robot.robot_model.set_base_xpos(xpos)
robot.robot_model.set_base_ori(rot)
else: # "single-arm-parallel" configuration setting
# Set up robots parallel to each other but offset from the center
for robot, offset in zip(self.robots, (-0.25, 0.25)):
xpos = robot.robot_model.base_xpos_offset["empty"]
xpos = np.array(xpos) + np.array((0, offset, 0))
robot.robot_model.set_base_xpos(xpos)
# Add arena and robot
mujoco_arena = EmptyArena()
# Arena always gets set to zero origin
mujoco_arena.set_origin([0, 0, 0])
# Modify default agentview camera
mujoco_arena.set_camera(camera_name="agentview",
pos=[
1.0666432116509934, 1.4903257668114777e-08,
2.0563394967349096
],
quat=[
0.6530979871749878, 0.27104058861732483,
0.27104055881500244, 0.6530978679656982
])
# initialize objects of interest
self.hole = PlateWithHoleObject(name="hole")
tex_attrib = {
"type": "cube",
}
mat_attrib = {
"texrepeat": "1 1",
"specular": "0.4",
"shininess": "0.1",
}
greenwood = CustomMaterial(
texture="WoodGreen",
tex_name="greenwood",
mat_name="greenwood_mat",
tex_attrib=tex_attrib,
mat_attrib=mat_attrib,
)
self.peg = CylinderObject(
name="peg",
size_min=(self.peg_radius[0], self.peg_length),
size_max=(self.peg_radius[1], self.peg_length),
material=greenwood,
rgba=[0, 1, 0, 1],
joints=None,
)
# Load hole object
hole_obj = self.hole.get_obj()
hole_obj.set("quat", "0 0 0.707 0.707")
hole_obj.set("pos", "0.11 0 0.17")
# Load peg object
peg_obj = self.peg.get_obj()
peg_obj.set("pos", array_to_string((0, 0, self.peg_length)))
# Append appropriate objects to arms
if self.env_configuration == "bimanual":
r_eef, l_eef = [
self.robots[0].robot_model.eef_name[arm]
for arm in self.robots[0].arms
]
r_model, l_model = [
self.robots[0].robot_model, self.robots[0].robot_model
]
else:
r_eef, l_eef = [
robot.robot_model.eef_name for robot in self.robots
]
r_model, l_model = [
self.robots[0].robot_model, self.robots[1].robot_model
]
r_body = find_elements(root=r_model.worldbody,
tags="body",
attribs={"name": r_eef},
return_first=True)
l_body = find_elements(root=l_model.worldbody,
tags="body",
attribs={"name": l_eef},
return_first=True)
r_body.append(peg_obj)
l_body.append(hole_obj)
# task includes arena, robot, and objects of interest
# We don't add peg and hole directly since they were already appended to the robots
self.model = ManipulationTask(
mujoco_arena=mujoco_arena,
mujoco_robots=[robot.robot_model for robot in self.robots],
)
# Make sure to add relevant assets from peg and hole objects
self.model.merge_assets(self.hole)
self.model.merge_assets(self.peg)
def _setup_references(self):
"""
Sets up references to important components. A reference is typically an
index or a list of indices that point to the corresponding elements
in a flatten array, which is how MuJoCo stores physical simulation data.
"""
super()._setup_references()
# Additional object references from this env
self.hole_body_id = self.sim.model.body_name2id(self.hole.root_body)
self.peg_body_id = self.sim.model.body_name2id(self.peg.root_body)
def _setup_observables(self):
"""
Sets up observables to be used for this environment. Creates object-based observables if enabled
Returns:
OrderedDict: Dictionary mapping observable names to its corresponding Observable object
"""
observables = super()._setup_observables()
# low-level object information
if self.use_object_obs:
# Get robot prefix and define observables modality
if self.env_configuration == "bimanual":
pf0 = self.robots[0].robot_model.naming_prefix + "right_"
pf1 = self.robots[0].robot_model.naming_prefix + "left_"
else:
pf0 = self.robots[0].robot_model.naming_prefix
pf1 = self.robots[1].robot_model.naming_prefix
modality = "object"
# position and rotation of peg and hole
@sensor(modality=modality)
def hole_pos(obs_cache):
return np.array(self.sim.data.body_xpos[self.hole_body_id])
@sensor(modality=modality)
def hole_quat(obs_cache):
return T.convert_quat(
self.sim.data.body_xquat[self.hole_body_id], to="xyzw")
@sensor(modality=modality)
def peg_to_hole(obs_cache):
return obs_cache["hole_pos"] - np.array(self.sim.data.body_xpos[self.peg_body_id]) if \
"hole_pos" in obs_cache else np.zeros(3)
@sensor(modality=modality)
def peg_quat(obs_cache):
return T.convert_quat(
self.sim.data.body_xquat[self.peg_body_id], to="xyzw")
# Relative orientation parameters
@sensor(modality=modality)
def angle(obs_cache):
t, d, cos = self._compute_orientation()
obs_cache["t"] = t
obs_cache["d"] = d
return cos
@sensor(modality=modality)
def t(obs_cache):
return obs_cache["t"] if "t" in obs_cache else 0.0
@sensor(modality=modality)
def d(obs_cache):
return obs_cache["d"] if "d" in obs_cache else 0.0
sensors = [hole_pos, hole_quat, peg_to_hole, peg_quat, angle, t, d]
names = [s.__name__ for s in sensors]
# Create observables
for name, s in zip(names, sensors):
observables[name] = Observable(
name=name,
sensor=s,
sampling_rate=self.control_freq,
)
return observables
def _reset_internal(self):
"""
Resets simulation internal configurations.
"""
super()._reset_internal()
def _check_success(self):
"""
Check if peg is successfully aligned and placed within the hole
Returns:
bool: True if peg is placed in hole correctly
"""
t, d, cos = self._compute_orientation()
return d < 0.06 and -0.12 <= t <= 0.14 and cos > 0.95
def _compute_orientation(self):
"""
Helper function to return the relative positions between the hole and the peg.
In particular, the intersection of the line defined by the peg and the plane
defined by the hole is computed; the parallel distance, perpendicular distance,
and angle are returned.
Returns:
3-tuple:
- (float): parallel distance
- (float): perpendicular distance
- (float): angle
"""
peg_mat = self.sim.data.body_xmat[self.peg_body_id]
peg_mat.shape = (3, 3)
peg_pos = self.sim.data.body_xpos[self.peg_body_id]
hole_pos = self.sim.data.body_xpos[self.hole_body_id]
hole_mat = self.sim.data.body_xmat[self.hole_body_id]
hole_mat.shape = (3, 3)
v = peg_mat @ np.array([0, 0, 1])
v = v / np.linalg.norm(v)
center = hole_pos + hole_mat @ np.array([0.1, 0, 0])
t = (center - peg_pos) @ v / (np.linalg.norm(v)**2)
d = np.linalg.norm(np.cross(v, peg_pos - center)) / np.linalg.norm(v)
hole_normal = hole_mat @ np.array([0, 0, 1])
return (
t,
d,
abs(
np.dot(hole_normal, v) / np.linalg.norm(hole_normal) /
np.linalg.norm(v)),
)
def _peg_pose_in_hole_frame(self):
"""
A helper function that takes in a named data field and returns the pose of that
object in the base frame.
Returns:
np.array: (4,4) matrix corresponding to the pose of the peg in the hole frame
"""
# World frame
peg_pos_in_world = self.sim.data.get_body_xpos(self.peg.root_body)
peg_rot_in_world = self.sim.data.get_body_xmat(
self.peg.root_body).reshape((3, 3))
peg_pose_in_world = T.make_pose(peg_pos_in_world, peg_rot_in_world)
# World frame
hole_pos_in_world = self.sim.data.get_body_xpos(self.hole.root_body)
hole_rot_in_world = self.sim.data.get_body_xmat(
self.hole.root_body).reshape((3, 3))
hole_pose_in_world = T.make_pose(hole_pos_in_world, hole_rot_in_world)
world_pose_in_hole = T.pose_inv(hole_pose_in_world)
peg_pose_in_hole = T.pose_in_A_to_pose_in_B(peg_pose_in_world,
world_pose_in_hole)
return peg_pose_in_hole