def reset_from_xml_string(self, xml_string):
"""Reloads the environment from an XML description of the environment."""
# if there is an active viewer window, destroy it
self.close()
# load model from xml
self.mjpy_model = load_model_from_xml(xml_string)
self.sim = MjSim(self.mjpy_model)
self.initialize_time(self.control_freq)
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0
)
self.viewer.viewer.vopt.geomgroup[1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
elif self.has_offscreen_renderer:
render_context = MjRenderContextOffscreen(self.sim)
render_context.vopt.geomgroup[0] = 1 if self.render_collision_mesh else 0
render_context.vopt.geomgroup[1] = 1 if self.render_visual_mesh else 0
self.sim.add_render_context(render_context)
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
# necessary to refresh MjData
self.sim.forward()
def _reset_internal(self):
"""Resets simulation internal configurations."""
# create visualization screen or renderer
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (1 if self.render_collision_mesh else 0)
self.viewer.viewer.vopt.geomgroup[1] = (1 if self.render_visual_mesh else 0)
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
# make sure mujoco-py doesn't block rendering frames
# (see https://github.com/StanfordVL/robosuite/issues/39)
self.viewer.viewer._render_every_frame = True
# Set the camera angle for viewing
if self.render_camera is not None:
self.viewer.set_camera(camera_id=self.sim.model.camera_name2id(self.render_camera))
elif self.has_offscreen_renderer:
if self.sim._render_context_offscreen is None:
render_context = MjRenderContextOffscreen(self.sim, device_id=self.render_gpu_device_id)
self.sim.add_render_context(render_context)
self.sim._render_context_offscreen.vopt.geomgroup[0] = (1 if self.render_collision_mesh else 0)
self.sim._render_context_offscreen.vopt.geomgroup[1] = (1 if self.render_visual_mesh else 0)
# additional housekeeping
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
def _reset_internal(self):
"""Resets simulation internal configurations."""
# instantiate simulation from MJCF model
self._load_model()
self.mjpy_model = self.model.get_model(mode="mujoco_py")
self.sim = MjSim(self.mjpy_model)
self.initialize_time(self.control_freq)
# create visualization screen or renderer
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.viewer.viewer.vopt.geomgroup[
1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
elif self.has_offscreen_renderer:
if self.sim._render_context_offscreen is None:
render_context = MjRenderContextOffscreen(self.sim)
self.sim.add_render_context(render_context)
self.sim._render_context_offscreen.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.sim._render_context_offscreen.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0)
# additional housekeeping
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
def init_viewer(self):
print('init_viewer', self.viewer)
if self.has_offscreen_renderer:
print(
'WARNING from GZZ: robosuite has a bug on simutaneously rendering for both offscreen (like camera obs) and onscreen (X window), and will give (at least) onscreen `black screen of death` after a reset. Please check their issue: https://github.com/StanfordVL/robosuite/issues/25 . I failed to help them debug on this.'
)
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (1 if self.render_collision_mesh
else 0)
self.viewer.viewer.vopt.geomgroup[
1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
def render(self):
"""
Fix Robosuite render method, so that Viewer is only spawned on render
"""
if self._viewer is None:
self._viewer = MujocoPyRenderer(self.sim)
self._viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self._viewer.viewer.vopt.geomgroup[
1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self._viewer.viewer._hide_overlay = True
self._viewer._render_every_frame = True
time.sleep(1 / self.control_freq)
self._viewer.render()
class MujocoEnv(metaclass=EnvMeta):
"""Initializes a Mujoco Environment."""
def __init__(
self,
has_renderer=False,
has_offscreen_renderer=True,
render_collision_mesh=False,
render_visual_mesh=True,
control_freq=10,
horizon=1000,
ignore_done=False,
use_camera_obs=False,
camera_name="frontview",
camera_height=256,
camera_width=256,
camera_depth=False,
):
"""
Args:
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_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 simulated 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).
use_camera_obs (bool): if True, every observation includes a
rendered image.
camera_name (str): name of camera to be rendered. Must be
set if @use_camera_obs is True.
camera_height (int): height of camera frame.
camera_width (int): width of camera frame.
camera_depth (bool): True if rendering RGB-D, and RGB otherwise.
"""
self.has_renderer = has_renderer
self.has_offscreen_renderer = has_offscreen_renderer
self.render_collision_mesh = render_collision_mesh
self.render_visual_mesh = render_visual_mesh
self.control_freq = control_freq
self.horizon = horizon
self.ignore_done = ignore_done
self.viewer = None
self.model = None
# settings for camera observations
self.use_camera_obs = use_camera_obs
if self.use_camera_obs and not self.has_offscreen_renderer:
raise ValueError("Camera observations require an offscreen renderer.")
self.camera_name = camera_name
if self.use_camera_obs and self.camera_name is None:
raise ValueError("Must specify camera name when using camera obs")
self.camera_height = camera_height
self.camera_width = camera_width
self.camera_depth = camera_depth
self._reset_internal()
def initialize_time(self, control_freq):
"""
Initializes the time constants used for simulation.
"""
self.cur_time = 0
self.model_timestep = self.sim.model.opt.timestep
if self.model_timestep <= 0:
raise XMLError("xml model defined non-positive time step")
self.control_freq = control_freq
if control_freq <= 0:
raise SimulationError(
"control frequency {} is invalid".format(control_freq)
)
self.control_timestep = 1. / control_freq
def _load_model(self):
"""Loads an xml model, puts it in self.model"""
pass
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.
"""
pass
def reset(self):
"""Resets simulation."""
# TODO(yukez): investigate black screen of death
# if there is an active viewer window, destroy it
self._destroy_viewer()
self._reset_internal()
self.sim.forward()
return self._get_observation()
def _reset_internal(self):
"""Resets simulation internal configurations."""
# instantiate simulation from MJCF model
self._load_model()
self.mjpy_model = self.model.get_model(mode="mujoco_py")
self.sim = MjSim(self.mjpy_model)
self.initialize_time(self.control_freq)
# create visualization screen or renderer
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0
)
self.viewer.viewer.vopt.geomgroup[1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
elif self.has_offscreen_renderer:
if self.sim._render_context_offscreen is None:
render_context = MjRenderContextOffscreen(self.sim)
self.sim.add_render_context(render_context)
self.sim._render_context_offscreen.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0
)
self.sim._render_context_offscreen.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0
)
# additional housekeeping
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
def _get_observation(self):
"""Returns an OrderedDict containing observations [(name_string, np.array), ...]."""
return OrderedDict()
def step(self, action):
"""Takes a step in simulation with control command @action."""
if self.done:
raise ValueError("executing action in terminated episode")
self.timestep += 1
self._pre_action(action)
end_time = self.cur_time + self.control_timestep
while self.cur_time < end_time:
self.sim.step()
self.cur_time += self.model_timestep
reward, done, info = self._post_action(action)
return self._get_observation(), reward, done, info
def _pre_action(self, action):
"""Do any preprocessing before taking an action."""
self.sim.data.ctrl[:] = action
def _post_action(self, action):
"""Do any housekeeping after taking an action."""
reward = self.reward(action)
# done if number of elapsed timesteps is greater than horizon
self.done = (self.timestep >= self.horizon) and not self.ignore_done
return reward, self.done, {}
def reward(self, action):
"""Reward should be a function of state and action."""
return 0
def render(self):
"""
Renders to an on-screen window.
"""
self.viewer.render()
def observation_spec(self):
"""
Returns an observation as observation specification.
An alternative design is to return an OrderedDict where the keys
are the observation names and the values are the shapes of observations.
We leave this alternative implementation commented out, as we find the
current design is easier to use in practice.
"""
observation = self._get_observation()
return observation
# observation_spec = OrderedDict()
# for k, v in observation.items():
# observation_spec[k] = v.shape
# return observation_spec
def action_spec(self):
"""
Action specification should be implemented in subclasses.
Action space is represented by a tuple of (low, high), which are two numpy
vectors that specify the min/max action limits per dimension.
"""
raise NotImplementedError
def reset_from_xml_string(self, xml_string):
"""Reloads the environment from an XML description of the environment."""
# if there is an active viewer window, destroy it
self.close()
# load model from xml
self.mjpy_model = load_model_from_xml(xml_string)
self.sim = MjSim(self.mjpy_model)
self.initialize_time(self.control_freq)
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0
)
self.viewer.viewer.vopt.geomgroup[1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
elif self.has_offscreen_renderer:
render_context = MjRenderContextOffscreen(self.sim)
render_context.vopt.geomgroup[0] = 1 if self.render_collision_mesh else 0
render_context.vopt.geomgroup[1] = 1 if self.render_visual_mesh else 0
self.sim.add_render_context(render_context)
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
# necessary to refresh MjData
self.sim.forward()
def find_contacts(self, geoms_1, geoms_2):
"""
Finds contact between two geom groups.
Args:
geoms_1: a list of geom names (string)
geoms_2: another list of geom names (string)
Returns:
iterator of all contacts between @geoms_1 and @geoms_2
"""
for contact in self.sim.data.contact[0 : self.sim.data.ncon]:
# check contact geom in geoms
c1_in_g1 = self.sim.model.geom_id2name(contact.geom1) in geoms_1
c2_in_g2 = self.sim.model.geom_id2name(contact.geom2) in geoms_2
# check contact geom in geoms (flipped)
c2_in_g1 = self.sim.model.geom_id2name(contact.geom2) in geoms_1
c1_in_g2 = self.sim.model.geom_id2name(contact.geom1) in geoms_2
if (c1_in_g1 and c2_in_g2) or (c1_in_g2 and c2_in_g1):
yield contact
def _check_contact(self):
"""Returns True if gripper is in contact with an object."""
return False
def _check_success(self):
"""
Returns True if task has been completed.
"""
return False
def _destroy_viewer(self):
# if there is an active viewer window, destroy it
if self.viewer is not None:
self.viewer.close() # change this to viewer.finish()?
self.viewer = None
def close(self):
"""Do any cleanup necessary here."""
self._destroy_viewer()
class MujocoEnv(metaclass=EnvMeta):
"""
Initializes a Mujoco Environment.
Args:
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 simulated 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
Raises:
ValueError: [Invalid renderer selection]
"""
def __init__(self,
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):
# First, verify that both the on- and off-screen renderers are not being used simultaneously
if has_renderer is True and has_offscreen_renderer is True:
raise ValueError(
"the onscreen and offscreen renderers cannot be used simultaneously."
)
# Rendering-specific attributes
self.has_renderer = has_renderer
self.has_offscreen_renderer = has_offscreen_renderer
self.render_camera = render_camera
self.render_collision_mesh = render_collision_mesh
self.render_visual_mesh = render_visual_mesh
self.render_gpu_device_id = render_gpu_device_id
self.viewer = None
# Simulation-specific attributes
self._observables = {} # Maps observable names to observable objects
self._obs_cache = {
} # Maps observable names to pre-/partially-computed observable values
self.control_freq = control_freq
self.horizon = horizon
self.ignore_done = ignore_done
self.hard_reset = hard_reset
self._model_postprocessor = None # Function to post-process model after load_model() call
self.model = None
self.cur_time = None
self.model_timestep = None
self.control_timestep = None
self.deterministic_reset = False # Whether to add randomized resetting of objects / robot joints
# Load the model
self._load_model()
# Post-process model
self._postprocess_model()
# Initialize the simulation
self._initialize_sim()
# Run all further internal (re-)initialization required
self._reset_internal()
# Load observables
self._observables = self._setup_observables()
def initialize_time(self, control_freq):
"""
Initializes the time constants used for simulation.
Args:
control_freq (float): Hz rate to run control loop at within the simulation
"""
self.cur_time = 0
self.model_timestep = macros.SIMULATION_TIMESTEP
if self.model_timestep <= 0:
raise ValueError("Invalid simulation timestep defined!")
self.control_freq = control_freq
if control_freq <= 0:
raise SimulationError(
"Control frequency {} is invalid".format(control_freq))
self.control_timestep = 1. / control_freq
def set_model_postprocessor(self, postprocessor):
"""
Sets the post-processor function that self.model will be passed to after load_model() is called during resets.
Args:
postprocessor (None or function): If set, postprocessing method should take in a Task-based instance and
return no arguments.
"""
self._model_postprocessor = postprocessor
def _load_model(self):
"""Loads an xml model, puts it in self.model"""
pass
def _postprocess_model(self):
"""
Post-processes model after load_model() call. Useful for external objects (e.g.: wrappers) to
be able to modify the sim model before it is actually loaded into the simulation
"""
if self._model_postprocessor is not None:
self._model_postprocessor(self.model)
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.
"""
pass
def _setup_observables(self):
"""
Sets up observables to be used for this environment.
Returns:
OrderedDict: Dictionary mapping observable names to its corresponding Observable object
"""
return OrderedDict()
def _initialize_sim(self, xml_string=None):
"""
Creates a MjSim object and stores it in self.sim. If @xml_string is specified, the MjSim object will be created
from the specified xml_string. Else, it will pull from self.model to instantiate the simulation
Args:
xml_string (str): If specified, creates MjSim object from this filepath
"""
# if we have an xml string, use that to create the sim. Otherwise, use the local model
self.mjpy_model = load_model_from_xml(
xml_string) if xml_string else self.model.get_model(
mode="mujoco_py")
# Create the simulation instance and run a single step to make sure changes have propagated through sim state
self.sim = MjSim(self.mjpy_model)
self.sim.forward()
# Setup sim time based on control frequency
self.initialize_time(self.control_freq)
def reset(self):
"""
Resets simulation.
Returns:
OrderedDict: Environment observation space after reset occurs
"""
# TODO(yukez): investigate black screen of death
# Use hard reset if requested
if self.hard_reset and not self.deterministic_reset:
self._destroy_viewer()
self._load_model()
self._postprocess_model()
self._initialize_sim()
# Else, we only reset the sim internally
else:
self.sim.reset()
# Reset necessary robosuite-centric variables
self._reset_internal()
self.sim.forward()
# Setup observables, reloading if
self._obs_cache = {}
if self.hard_reset:
# If we're using hard reset, must re-update sensor object references
_observables = self._setup_observables()
for obs_name, obs in _observables.items():
self.modify_observable(observable_name=obs_name,
attribute="sensor",
modifier=obs._sensor)
# Make sure that all sites are toggled OFF by default
self.visualize(
vis_settings={vis: False
for vis in self._visualizations})
# Return new observations
return self._get_observations()
def _reset_internal(self):
"""Resets simulation internal configurations."""
# create visualization screen or renderer
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.viewer.viewer.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0)
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
# make sure mujoco-py doesn't block rendering frames
# (see https://github.com/StanfordVL/robosuite/issues/39)
self.viewer.viewer._render_every_frame = True
# Set the camera angle for viewing
if self.render_camera is not None:
self.viewer.set_camera(camera_id=self.sim.model.camera_name2id(
self.render_camera))
elif self.has_offscreen_renderer:
if self.sim._render_context_offscreen is None:
render_context = MjRenderContextOffscreen(
self.sim, device_id=self.render_gpu_device_id)
self.sim.add_render_context(render_context)
self.sim._render_context_offscreen.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.sim._render_context_offscreen.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0)
# additional housekeeping
self.sim_state_initial = self.sim.get_state()
self._setup_references()
self.cur_time = 0
self.timestep = 0
self.done = False
# Empty observation cache and reset all observables
self._obs_cache = {}
for observable in self._observables.values():
observable.reset()
def _update_observables(self):
"""
Updates all observables in this environment
"""
for observable in self._observables.values():
observable.update(timestep=self.model_timestep,
obs_cache=self._obs_cache)
def _get_observations(self):
"""
Grabs observations from the environment.
Returns:
OrderedDict: OrderedDict containing observations [(name_string, np.array), ...]
"""
observations = OrderedDict()
obs_by_modality = OrderedDict()
# Loop through all observables and grab their current observation
for obs_name, observable in self._observables.items():
if observable.is_enabled() and observable.is_active():
obs = observable.obs
observations[obs_name] = obs
modality = observable.modality + "-state"
if modality not in obs_by_modality:
obs_by_modality[modality] = []
# Make sure all observations are numpy arrays so we can concatenate them
array_obs = [obs] if type(obs) in {int, float
} or not obs.shape else obs
obs_by_modality[modality].append(np.array(array_obs))
# Add in modality observations
for modality, obs in obs_by_modality.items():
# To save memory, we only concatenate the image observations if explicitly requested
if modality == "image-state" and not macros.CONCATENATE_IMAGES:
continue
observations[modality] = np.concatenate(obs, axis=-1)
return observations
def step(self, action):
"""
Takes a step in simulation with control command @action.
Args:
action (np.array): Action to execute within the environment
Returns:
4-tuple:
- (OrderedDict) observations from the environment
- (float) reward from the environment
- (bool) whether the current episode is completed or not
- (dict) misc information
Raises:
ValueError: [Steps past episode termination]
"""
if self.done:
raise ValueError("executing action in terminated episode")
self.timestep += 1
# Since the env.step frequency is slower than the mjsim timestep frequency, the internal controller will output
# multiple torque commands in between new high level action commands. Therefore, we need to denote via
# 'policy_step' whether the current step we're taking is simply an internal update of the controller,
# or an actual policy update
policy_step = True
# Loop through the simulation at the model timestep rate until we're ready to take the next policy step
# (as defined by the control frequency specified at the environment level)
for i in range(int(self.control_timestep / self.model_timestep)):
self.sim.forward()
self._pre_action(action, policy_step)
self.sim.step()
self._update_observables()
policy_step = False
# Note: this is done all at once to avoid floating point inaccuracies
self.cur_time += self.control_timestep
reward, done, info = self._post_action(action)
return self._get_observations(), reward, done, info
def _pre_action(self, action, policy_step=False):
"""
Do any preprocessing before taking an action.
Args:
action (np.array): Action to execute within the environment
policy_step (bool): Whether this current loop is an actual policy step or internal sim update step
"""
self.sim.data.ctrl[:] = action
def _post_action(self, action):
"""
Do any housekeeping after taking an action.
Args:
action (np.array): Action to execute within the environment
Returns:
3-tuple:
- (float) reward from the environment
- (bool) whether the current episode is completed or not
- (dict) empty dict to be filled with information by subclassed method
"""
reward = self.reward(action)
# done if number of elapsed timesteps is greater than horizon
self.done = (self.timestep >= self.horizon) and not self.ignore_done
return reward, self.done, {}
def reward(self, action):
"""
Reward should be a function of state and action
Args:
action (np.array): Action to execute within the environment
Returns:
float: Reward from environment
"""
raise NotImplementedError
def render(self):
"""
Renders to an on-screen window.
"""
self.viewer.render()
def observation_spec(self):
"""
Returns an observation as observation specification.
An alternative design is to return an OrderedDict where the keys
are the observation names and the values are the shapes of observations.
We leave this alternative implementation commented out, as we find the
current design is easier to use in practice.
Returns:
OrderedDict: Observations from the environment
"""
observation = self._get_observations()
return observation
def clear_objects(self, object_names):
"""
Clears objects with the name @object_names out of the task space. This is useful
for supporting task modes with single types of objects, as in
@self.single_object_mode without changing the model definition.
Args:
object_names (str or list of str): Name of object(s) to remove from the task workspace
"""
object_names = {object_names
} if type(object_names) is str else set(object_names)
for obj in self.model.mujoco_objects:
if obj.name in object_names:
self.sim.data.set_joint_qpos(
obj.joints[0], np.array((10, 10, 10, 1, 0, 0, 0)))
def visualize(self, vis_settings):
"""
Do any needed visualization here
Args:
vis_settings (dict): Visualization keywords mapped to T/F, determining whether that specific
component should be visualized. Should have "env" keyword as well as any other relevant
options specified.
"""
# Set visuals for environment objects
for obj in self.model.mujoco_objects:
obj.set_sites_visibility(sim=self.sim, visible=vis_settings["env"])
def reset_from_xml_string(self, xml_string):
"""
Reloads the environment from an XML description of the environment.
Args:
xml_string (str): Filepath to the xml file that will be loaded directly into the sim
"""
# if there is an active viewer window, destroy it
self.close()
# Since we are reloading from an xml_string, we are deterministically resetting
self.deterministic_reset = True
# initialize sim from xml
self._initialize_sim(xml_string=xml_string)
# Now reset as normal
self.reset()
# Turn off deterministic reset
self.deterministic_reset = False
def check_contact(self, geoms_1, geoms_2=None):
"""
Finds contact between two geom groups.
Args:
geoms_1 (str or list of str or MujocoModel): an individual geom name or list of geom names or a model. If
a MujocoModel is specified, the geoms checked will be its contact_geoms
geoms_2 (str or list of str or MujocoModel or None): another individual geom name or list of geom names.
If a MujocoModel is specified, the geoms checked will be its contact_geoms. If None, will check
any collision with @geoms_1 to any other geom in the environment
Returns:
bool: True if any geom in @geoms_1 is in contact with any geom in @geoms_2.
"""
# Check if either geoms_1 or geoms_2 is a string, convert to list if so
if type(geoms_1) is str:
geoms_1 = [geoms_1]
elif isinstance(geoms_1, MujocoModel):
geoms_1 = geoms_1.contact_geoms
if type(geoms_2) is str:
geoms_2 = [geoms_2]
elif isinstance(geoms_2, MujocoModel):
geoms_2 = geoms_2.contact_geoms
for contact in self.sim.data.contact[:self.sim.data.ncon]:
# check contact geom in geoms
c1_in_g1 = self.sim.model.geom_id2name(contact.geom1) in geoms_1
c2_in_g2 = self.sim.model.geom_id2name(
contact.geom2) in geoms_2 if geoms_2 is not None else True
# check contact geom in geoms (flipped)
c2_in_g1 = self.sim.model.geom_id2name(contact.geom2) in geoms_1
c1_in_g2 = self.sim.model.geom_id2name(
contact.geom1) in geoms_2 if geoms_2 is not None else True
if (c1_in_g1 and c2_in_g2) or (c1_in_g2 and c2_in_g1):
return True
return False
def get_contacts(self, model):
"""
Checks for any contacts with @model (as defined by @model's contact_geoms) and returns the set of
geom names currently in contact with that model (excluding the geoms that are part of the model itself).
Args:
model (MujocoModel): Model to check contacts for.
Returns:
set: Unique geoms that are actively in contact with this model.
Raises:
AssertionError: [Invalid input type]
"""
# Make sure model is MujocoModel type
assert isinstance(model, MujocoModel), \
"Inputted model must be of type MujocoModel; got type {} instead!".format(type(model))
contact_set = set()
for contact in self.sim.data.contact[:self.sim.data.ncon]:
# check contact geom in geoms; add to contact set if match is found
g1, g2 = self.sim.model.geom_id2name(
contact.geom1), self.sim.model.geom_id2name(contact.geom2)
if g1 in model.contact_geoms and g2 not in model.contact_geoms:
contact_set.add(g2)
elif g2 in model.contact_geoms and g1 not in model.contact_geoms:
contact_set.add(g1)
return contact_set
def add_observable(self, observable):
"""
Adds an observable to this environment.
Args:
observable (Observable): Observable instance.
"""
assert observable.name not in self._observables,\
"Observable name {} is already associated with an existing observable! Use modify_observable(...) " \
"to modify a pre-existing observable.".format(observable.name)
self._observables[observable.name] = observable
def modify_observable(self, observable_name, attribute, modifier):
"""
Modifies observable with associated name @observable_name, replacing the given @attribute with @modifier.
Args:
observable_name (str): Observable to modify
attribute (str): Observable attribute to modify.
Options are {`'sensor'`, `'corrupter'`, `'delayer'`, `'sampling_rate'`, `'enabled'`, `'active'`}
modifier (any): New function / value to replace with for observable. If a function, new signature should
match the function being replaced.
"""
# Find the observable
assert observable_name in self._observables, "No valid observable with name {} found. Options are: {}".\
format(observable_name, self.observation_names)
obs = self._observables[observable_name]
# replace attribute accordingly
if attribute == "sensor":
obs.set_sensor(modifier)
elif attribute == "corrupter":
obs.set_corrupter(modifier)
elif attribute == "delayer":
obs.set_delayer(modifier)
elif attribute == "sampling_rate":
obs.set_sampling_rate(modifier)
elif attribute == "enabled":
obs.set_enabled(modifier)
elif attribute == "active":
obs.set_active(modifier)
else:
# Invalid attribute specified
raise ValueError(
"Invalid observable attribute specified. Requested: {}, valid options are {}"
.format(
attribute, {
"sensor", "corrupter", "delayer", "sampling_rate",
"enabled", "active"
}))
def _check_success(self):
"""
Checks if the task has been completed. Should be implemented by subclasses
Returns:
bool: True if the task has been completed
"""
raise NotImplementedError
def _destroy_viewer(self):
"""
Destroys the current mujoco renderer instance if it exists
"""
# if there is an active viewer window, destroy it
if self.viewer is not None:
self.viewer.close() # change this to viewer.finish()?
self.viewer = None
def close(self):
"""Do any cleanup necessary here."""
self._destroy_viewer()
@property
def observation_modalities(self):
"""
Modalities for this environment's observations
Returns:
set: All observation modalities
"""
return set(
[observable.modality for observable in self._observables.values()])
@property
def observation_names(self):
"""
Grabs all names for this environment's observables
Returns:
set: All observation names
"""
return set(self._observables.keys())
@property
def enabled_observables(self):
"""
Grabs all names of enabled observables for this environment. An observable is considered enabled if its values
are being continually computed / updated at each simulation timestep.
Returns:
set: All enabled observation names
"""
return set([
name for name, observable in self._observables.items()
if observable.is_enabled()
])
@property
def active_observables(self):
"""
Grabs all names of active observables for this environment. An observable is considered active if its value is
being returned in the observation dict from _get_observations() call or from the step() call (assuming this
observable is enabled).
Returns:
set: All active observation names
"""
return set([
name for name, observable in self._observables.items()
if observable.is_active()
])
@property
def _visualizations(self):
"""
Visualization keywords for this environment
Returns:
set: All components that can be individually visualized for this environment
"""
return {"env"}
@property
def action_spec(self):
"""
Action specification should be implemented in subclasses.
Action space is represented by a tuple of (low, high), which are two numpy
vectors that specify the min/max action limits per dimension.
"""
raise NotImplementedError
@property
def action_dim(self):
"""
Size of the action space
Returns:
int: Action space dimension
"""
raise NotImplementedError
class GoalBasedSawyerLift(SawyerLift, gym.GoalEnv):
"""
Wraps SawyerLift in a GoalEnv.
"""
def __init__(self, test=False, use_dense=True, objgrip=False):
super().__init__(use_camera_obs=False,
has_renderer=False,
use_indicator_object=True,
has_offscreen_renderer=False,
horizon=50,
ignore_done=True)
self._viewer = None
self.action_space = gym.spaces.Box(*self.action_spec)
self._init_eef = np.array([0.58218024, -0.01407946, 0.90169981])
o = self._get_observation()
self._obs_keys = [
'cube_pos', 'cube_quat', 'eef_pos', 'joint_pos', 'joint_vel',
'gripper_qpos', 'gripper_qvel'
]
observation_space = gym.spaces.Box(
-np.inf, np.inf, (sum([len(o[k]) for k in self._obs_keys]), ))
if objgrip:
self.ag = lambda obs: np.concatenate((obs[:3], obs[7:10]))
else:
self.ag = lambda obs: obs[:3]
if objgrip:
goal_space = gym.spaces.Box(-np.inf, np.inf, (6, ))
else:
goal_space = gym.spaces.Box(-np.inf, np.inf, (3, ))
self.observation_space = gym.spaces.Dict({
'observation': observation_space,
'desired_goal': goal_space,
'achieved_goal': goal_space
})
self.max_steps = 50
self.num_steps = 0
self.dist_threshold = 0.05
if objgrip:
self.dist_threshold = 0.07
self.test = test
self.use_dense = use_dense
self.objgrip = objgrip
self.goal = None
self._prev_state = None
def reset(self):
"""
Need to override reset because the original is awfully inefficient (it reloads Mujoco Sim entirely)
"""
self.num_steps = 0
# From base.py
self.sim.set_state(self.sim_state_initial)
self.initialize_time(self.control_freq)
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
# From sawyer.py
self.sim.data.qpos[
self._ref_joint_pos_indexes] = self.mujoco_robot.init_qpos
if self.has_gripper:
self.sim.data.qpos[
self._ref_gripper_joint_pos_indexes] = self.gripper.init_qpos
# From sawyer_lift.py
# reset positions of objects
# self.model.place_objects() <- doesn't actually work, because this operates on the model and not the sim...
init_pos = self.sim.data.get_joint_qpos('cube')
init_pos[:2] += np.random.uniform(-0.1, 0.1, 2)
self.sim.data.set_joint_qpos('cube', init_pos)
# reset goal position
self.goal = init_pos[:3] + np.array([0., 0., 0.12])
self.move_indicator(self.goal)
if self.objgrip:
self.goal = np.concatenate((self.goal, self.goal))
# reset joint positions
init_pos = np.array(
[-0.5538, -0.8208, 0.4155, 1.8409, -0.4955, 0.6482, 1.9628])
init_pos += np.random.randn(init_pos.shape[0]) * 0.02
self.sim.data.qpos[self._ref_joint_pos_indexes] = np.array(init_pos)
# And again from base.py
self.sim.forward()
obs = self._get_observation()
obs = np.concatenate([obs[k] for k in self._obs_keys])
ag = self.ag(obs)
self._prev_state = obs
obs = {
'observation': obs,
'achieved_goal': ag,
'desired_goal': self.goal,
}
return obs
def seed(self, seed=None):
np.random.seed(seed)
def compute_reward(self, ag, dg, info):
d = np.linalg.norm(ag - dg, axis=-1)
reward = -(d >= self.dist_threshold).astype(np.float32)
success = 1. + reward
failure = -1 * reward
reward -= d * success
ns = info['ns']
if len(ns.shape) == 1:
reward -= (
0.2 *
np.tanh(np.mean(np.abs(info['s'][7:17] - ns[7:17]), axis=-1)) *
success) # Penalize successes that move
else:
reward -= (0.2 * np.tanh(
np.mean(np.abs(info['s'][:, 7:17] - ns[:, 7:17]), axis=-1)) *
success) # Penalize successes that move
# add dense reward discourages agent from being far away from the cube
if self.use_dense:
if len(ns.shape) == 1:
reward -= np.linalg.norm(ns[:3] - ns[7:10]) * failure
else:
reward -= np.linalg.norm(ns[:, :3] - ns[:, 7:10],
axis=-1) * failure
return reward
def is_success(self, ag, dg):
d = np.linalg.norm(ag[:3] - dg[:3], axis=-1)
return d <= self.dist_threshold
def step(self, action):
obs, _, _, _ = super().step(action)
obs = np.concatenate([obs[k] for k in self._obs_keys])
ag = self.ag(obs)
eef = obs[7:10]
reward = self.compute_reward(ag, self.goal, {
's': obs,
'ns': self._prev_state
})
self._prev_state = obs
obs = {
'observation': obs,
'achieved_goal': ag,
'desired_goal': self.goal,
}
if not self.test:
info = {'is_success': self.is_success(ag, self.goal)}
elif self.test:
info = {'is_success': self._check_success()}
self.num_steps += 1
done = True if self.num_steps >= self.max_steps else False
if done: info['TimeLimit.truncated'] = True
return obs, reward, done, info
def render(self):
"""
Fix Robosuite render method, so that Viewer is only spawned on render
"""
if self._viewer is None:
self._viewer = MujocoPyRenderer(self.sim)
self._viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self._viewer.viewer.vopt.geomgroup[
1] = 1 if self.render_visual_mesh else 0
# hiding the overlay speeds up rendering significantly
self._viewer.viewer._hide_overlay = True
self._viewer.viewer._render_every_frame = True
time.sleep(1 / self.control_freq)
self._viewer.render()
class MujocoEnv(metaclass=EnvMeta):
"""
Initializes a Mujoco Environment.
Args:
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 simulated 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
Raises:
ValueError: [Invalid renderer selection]
"""
def __init__(self,
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):
# First, verify that both the on- and off-screen renderers are not being used simultaneously
if has_renderer is True and has_offscreen_renderer is True:
raise ValueError(
"the onscreen and offscreen renderers cannot be used simultaneously."
)
# Rendering-specific attributes
self.has_renderer = has_renderer
self.has_offscreen_renderer = has_offscreen_renderer
self.render_camera = render_camera
self.render_collision_mesh = render_collision_mesh
self.render_visual_mesh = render_visual_mesh
self.viewer = None
# Simulation-specific attributes
self.control_freq = control_freq
self.horizon = horizon
self.ignore_done = ignore_done
self.hard_reset = hard_reset
self.model = None
self.cur_time = None
self.model_timestep = None
self.control_timestep = None
self.deterministic_reset = False # Whether to add randomized resetting of objects / robot joints
# Load the model
self._load_model()
# Initialize the simulation
self._initialize_sim()
# Run all further internal (re-)initialization required
self._reset_internal()
def initialize_time(self, control_freq):
"""
Initializes the time constants used for simulation.
Args:
control_freq (float): Hz rate to run control loop at within the simulation
"""
self.cur_time = 0
self.model_timestep = self.sim.model.opt.timestep
if self.model_timestep <= 0:
raise XMLError("xml model defined non-positive time step")
self.control_freq = control_freq
if control_freq <= 0:
raise SimulationError(
"control frequency {} is invalid".format(control_freq))
self.control_timestep = 1. / control_freq
def _load_model(self):
"""Loads an xml model, puts it in self.model"""
pass
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.
"""
pass
def _initialize_sim(self, xml_string=None):
"""
Creates a MjSim object and stores it in self.sim. If @xml_string is specified, the MjSim object will be created
from the specified xml_string. Else, it will pull from self.model to instantiate the simulation
Args:
xml_string (str): If specified, creates MjSim object from this filepath
"""
# if we have an xml string, use that to create the sim. Otherwise, use the local model
self.mjpy_model = load_model_from_xml(
xml_string) if xml_string else self.model.get_model(
mode="mujoco_py")
# Create the simulation instance and run a single step to make sure changes have propagated through sim state
self.sim = MjSim(self.mjpy_model)
self.sim.step()
# Setup sim time based on control frequency
self.initialize_time(self.control_freq)
def reset(self):
"""
Resets simulation.
Returns:
OrderedDict: Environment observation space after reset occurs
"""
# TODO(yukez): investigate black screen of death
# Use hard reset if requested
if self.hard_reset and not self.deterministic_reset:
self._destroy_viewer()
self._load_model()
self._initialize_sim()
# Else, we only reset the sim internally
else:
self.sim.reset()
# Reset necessary robosuite-centric variables
self._reset_internal()
self.sim.forward()
return self._get_observation()
def _reset_internal(self):
"""Resets simulation internal configurations."""
# create visualization screen or renderer
if self.has_renderer and self.viewer is None:
self.viewer = MujocoPyRenderer(self.sim)
self.viewer.viewer.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.viewer.viewer.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0)
# hiding the overlay speeds up rendering significantly
self.viewer.viewer._hide_overlay = True
# make sure mujoco-py doesn't block rendering frames
# (see https://github.com/StanfordVL/robosuite/issues/39)
self.viewer.viewer._render_every_frame = True
# Set the camera angle for viewing
if self.render_camera is not None:
self.viewer.set_camera(camera_id=self.sim.model.camera_name2id(
self.render_camera))
elif self.has_offscreen_renderer:
if self.sim._render_context_offscreen is None:
render_context = MjRenderContextOffscreen(self.sim)
self.sim.add_render_context(render_context)
self.sim._render_context_offscreen.vopt.geomgroup[0] = (
1 if self.render_collision_mesh else 0)
self.sim._render_context_offscreen.vopt.geomgroup[1] = (
1 if self.render_visual_mesh else 0)
# additional housekeeping
self.sim_state_initial = self.sim.get_state()
self._get_reference()
self.cur_time = 0
self.timestep = 0
self.done = False
def _get_observation(self):
"""
Grabs observations from the environment.
Returns:
OrderedDict: OrderedDict containing observations [(name_string, np.array), ...]
"""
return OrderedDict()
def step(self, action):
"""
Takes a step in simulation with control command @action.
Args:
action (np.array): Action to execute within the environment
Returns:
4-tuple:
- (OrderedDict) observations from the environment
- (float) reward from the environment
- (bool) whether the current episode is completed or not
- (dict) misc information
Raises:
ValueError: [Steps past episode termination]
"""
if self.done:
raise ValueError("executing action in terminated episode")
self.timestep += 1
# Since the env.step frequency is slower than the mjsim timestep frequency, the internal controller will output
# multiple torque commands in between new high level action commands. Therefore, we need to denote via
# 'policy_step' whether the current step we're taking is simply an internal update of the controller,
# or an actual policy update
policy_step = True
# Loop through the simulation at the model timestep rate until we're ready to take the next policy step
# (as defined by the control frequency specified at the environment level)
for i in range(int(self.control_timestep / self.model_timestep)):
self.sim.forward()
self._pre_action(action, policy_step)
self.sim.step()
policy_step = False
# Note: this is done all at once to avoid floating point inaccuracies
self.cur_time += self.control_timestep
reward, done, info = self._post_action(action)
return self._get_observation(), reward, done, info
def _pre_action(self, action, policy_step=False):
"""
Do any preprocessing before taking an action.
Args:
action (np.array): Action to execute within the environment
policy_step (bool): Whether this current loop is an actual policy step or internal sim update step
"""
self.sim.data.ctrl[:] = action
def _post_action(self, action):
"""
Do any housekeeping after taking an action.
Args:
action (np.array): Action to execute within the environment
Returns:
3-tuple:
- (float) reward from the environment
- (bool) whether the current episode is completed or not
- (dict) empty dict to be filled with information by subclassed method
"""
reward = self.reward(action)
# done if number of elapsed timesteps is greater than horizon
self.done = (self.timestep >= self.horizon) and not self.ignore_done
return reward, self.done, {}
def reward(self, action):
"""
Reward should be a function of state and action
Args:
action (np.array): Action to execute within the environment
Returns:
float: Reward from environment
"""
raise NotImplementedError
def render(self):
"""
Renders to an on-screen window.
"""
self.viewer.render()
def observation_spec(self):
"""
Returns an observation as observation specification.
An alternative design is to return an OrderedDict where the keys
are the observation names and the values are the shapes of observations.
We leave this alternative implementation commented out, as we find the
current design is easier to use in practice.
Returns:
OrderedDict: Observations from the environment
"""
observation = self._get_observation()
return observation
@property
def action_spec(self):
"""
Action specification should be implemented in subclasses.
Action space is represented by a tuple of (low, high), which are two numpy
vectors that specify the min/max action limits per dimension.
"""
raise NotImplementedError
@property
def action_dim(self):
"""
Size of the action space
Returns:
int: Action space dimension
"""
raise NotImplementedError
def reset_from_xml_string(self, xml_string):
"""
Reloads the environment from an XML description of the environment.
Args:
xml_string (str): Filepath to the xml file that will be loaded directly into the sim
"""
# if there is an active viewer window, destroy it
self.close()
# Since we are reloading from an xml_string, we are deterministically resetting
self.deterministic_reset = True
# initialize sim from xml
self._initialize_sim(xml_string=xml_string)
# Now reset as normal
self.reset()
# Turn off deterministic reset
self.deterministic_reset = False
def find_contacts(self, geoms_1, geoms_2):
"""
Finds contact between two geom groups.
Args:
geoms_1 (list of str): a list of geom names
geoms_2 (list of str): another list of geom names
Returns:
generator: iterator of all contacts between @geoms_1 and @geoms_2
"""
for contact in self.sim.data.contact[0:self.sim.data.ncon]:
# check contact geom in geoms
c1_in_g1 = self.sim.model.geom_id2name(contact.geom1) in geoms_1
c2_in_g2 = self.sim.model.geom_id2name(contact.geom2) in geoms_2
# check contact geom in geoms (flipped)
c2_in_g1 = self.sim.model.geom_id2name(contact.geom2) in geoms_1
c1_in_g2 = self.sim.model.geom_id2name(contact.geom1) in geoms_2
if (c1_in_g1 and c2_in_g2) or (c1_in_g2 and c2_in_g1):
yield contact
def _check_success(self):
"""
Checks if the task has been completed. Should be implemented by subclasses
Returns:
bool: True if the task has been completed
"""
raise NotImplementedError
def _destroy_viewer(self):
"""
Destroys the current mujoco renderer instance if it exists
"""
# if there is an active viewer window, destroy it
if self.viewer is not None:
self.viewer.close() # change this to viewer.finish()?
self.viewer = None
def close(self):
"""Do any cleanup necessary here."""
self._destroy_viewer()