# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause
import torch
from collections.abc import Sequence
from typing import Any
import carb
import omni.usd
from isaacsim.core.cloner import GridCloner
from isaacsim.core.prims import XFormPrim
from pxr import PhysxSchema
import isaaclab.sim as sim_utils
from isaaclab.assets import (
Articulation,
ArticulationCfg,
AssetBaseCfg,
DeformableObject,
DeformableObjectCfg,
RigidObject,
RigidObjectCfg,
RigidObjectCollection,
RigidObjectCollectionCfg,
)
from isaaclab.sensors import ContactSensorCfg, FrameTransformerCfg, SensorBase, SensorBaseCfg
from isaaclab.terrains import TerrainImporter, TerrainImporterCfg
from .interactive_scene_cfg import InteractiveSceneCfg
[docs]class InteractiveScene:
"""A scene that contains entities added to the simulation.
The interactive scene parses the :class:`InteractiveSceneCfg` class to create the scene.
Based on the specified number of environments, it clones the entities and groups them into different
categories (e.g., articulations, sensors, etc.).
Cloning can be performed in two ways:
* For tasks where all environments contain the same assets, a more performant cloning paradigm
can be used to allow for faster environment creation. This is specified by the ``replicate_physics`` flag.
.. code-block:: python
scene = InteractiveScene(cfg=InteractiveSceneCfg(replicate_physics=True))
* For tasks that require having separate assets in the environments, ``replicate_physics`` would have to
be set to False, which will add some costs to the overall startup time.
.. code-block:: python
scene = InteractiveScene(cfg=InteractiveSceneCfg(replicate_physics=False))
Each entity is registered to scene based on its name in the configuration class. For example, if the user
specifies a robot in the configuration class as follows:
.. code-block:: python
from isaaclab.scene import InteractiveSceneCfg
from isaaclab.utils import configclass
from isaaclab_assets.robots.anymal import ANYMAL_C_CFG
@configclass
class MySceneCfg(InteractiveSceneCfg):
robot = ANYMAL_C_CFG.replace(prim_path="{ENV_REGEX_NS}/Robot")
Then the robot can be accessed from the scene as follows:
.. code-block:: python
from isaaclab.scene import InteractiveScene
# create 128 environments
scene = InteractiveScene(cfg=MySceneCfg(num_envs=128))
# access the robot from the scene
robot = scene["robot"]
# access the robot based on its type
robot = scene.articulations["robot"]
If the :class:`InteractiveSceneCfg` class does not include asset entities, the cloning process
can still be triggered if assets were added to the stage outside of the :class:`InteractiveScene` class:
.. code-block:: python
scene = InteractiveScene(cfg=InteractiveSceneCfg(num_envs=128, replicate_physics=True))
scene.clone_environments()
.. note::
It is important to note that the scene only performs common operations on the entities. For example,
resetting the internal buffers, writing the buffers to the simulation and updating the buffers from the
simulation. The scene does not perform any task specific to the entity. For example, it does not apply
actions to the robot or compute observations from the robot. These tasks are handled by different
modules called "managers" in the framework. Please refer to the :mod:`isaaclab.managers` sub-package
for more details.
"""
[docs] def __init__(self, cfg: InteractiveSceneCfg):
"""Initializes the scene.
Args:
cfg: The configuration class for the scene.
"""
# check that the config is valid
cfg.validate()
# store inputs
self.cfg = cfg
# initialize scene elements
self._terrain = None
self._articulations = dict()
self._deformable_objects = dict()
self._rigid_objects = dict()
self._rigid_object_collections = dict()
self._sensors = dict()
self._extras = dict()
# obtain the current stage
self.stage = omni.usd.get_context().get_stage()
# physics scene path
self._physics_scene_path = None
# prepare cloner for environment replication
self.cloner = GridCloner(spacing=self.cfg.env_spacing)
self.cloner.define_base_env(self.env_ns)
self.env_prim_paths = self.cloner.generate_paths(f"{self.env_ns}/env", self.cfg.num_envs)
# create source prim
self.stage.DefinePrim(self.env_prim_paths[0], "Xform")
# when replicate_physics=False, we assume heterogeneous environments and clone the xforms first.
# this triggers per-object level cloning in the spawner.
if not self.cfg.replicate_physics:
# clone the env xform
env_origins = self.cloner.clone(
source_prim_path=self.env_prim_paths[0],
prim_paths=self.env_prim_paths,
replicate_physics=False,
copy_from_source=True,
enable_env_ids=self.cfg.filter_collisions, # this won't do anything because we are not replicating physics
)
self._default_env_origins = torch.tensor(env_origins, device=self.device, dtype=torch.float32)
else:
# otherwise, environment origins will be initialized during cloning at the end of environment creation
self._default_env_origins = None
self._global_prim_paths = list()
if self._is_scene_setup_from_cfg():
# add entities from config
self._add_entities_from_cfg()
# clone environments on a global scope if environment is homogeneous
if self.cfg.replicate_physics:
self.clone_environments(copy_from_source=False)
# replicate physics if we have more than one environment
# this is done to make scene initialization faster at play time
if self.cfg.replicate_physics and self.cfg.num_envs > 1:
self.cloner.replicate_physics(
source_prim_path=self.env_prim_paths[0],
prim_paths=self.env_prim_paths,
base_env_path=self.env_ns,
root_path=self.env_regex_ns.replace(".*", ""),
enable_env_ids=self.cfg.filter_collisions,
)
# since env_ids is only applicable when replicating physics, we have to fallback to the previous method
# to filter collisions if replicate_physics is not enabled
if not self.cfg.replicate_physics and self.cfg.filter_collisions:
self.filter_collisions(self._global_prim_paths)
[docs] def clone_environments(self, copy_from_source: bool = False):
"""Creates clones of the environment ``/World/envs/env_0``.
Args:
copy_from_source: (bool): If set to False, clones inherit from /World/envs/env_0 and mirror its changes.
If True, clones are independent copies of the source prim and won't reflect its changes (start-up time
may increase). Defaults to False.
"""
# check if user spawned different assets in individual environments
# this flag will be None if no multi asset is spawned
carb_settings_iface = carb.settings.get_settings()
has_multi_assets = carb_settings_iface.get("/isaaclab/spawn/multi_assets")
if has_multi_assets and self.cfg.replicate_physics:
omni.log.warn(
"Varying assets might have been spawned under different environments."
" However, the replicate physics flag is enabled in the 'InteractiveScene' configuration."
" This may adversely affect PhysX parsing. We recommend disabling this property."
)
# clone the environment
env_origins = self.cloner.clone(
source_prim_path=self.env_prim_paths[0],
prim_paths=self.env_prim_paths,
replicate_physics=self.cfg.replicate_physics,
copy_from_source=copy_from_source,
enable_env_ids=self.cfg.filter_collisions, # this automatically filters collisions between environments
)
# since env_ids is only applicable when replicating physics, we have to fallback to the previous method
# to filter collisions if replicate_physics is not enabled
if not self.cfg.replicate_physics and self.cfg.filter_collisions:
omni.log.warn(
"Collision filtering can only be automatically enabled when replicate_physics=True."
" Please call scene.filter_collisions(global_prim_paths) to filter collisions across environments."
)
# in case of heterogeneous cloning, the env origins is specified at init
if self._default_env_origins is None:
self._default_env_origins = torch.tensor(env_origins, device=self.device, dtype=torch.float32)
[docs] def filter_collisions(self, global_prim_paths: list[str] | None = None):
"""Filter environments collisions.
Disables collisions between the environments in ``/World/envs/env_.*`` and enables collisions with the prims
in global prim paths (e.g. ground plane).
Args:
global_prim_paths: A list of global prim paths to enable collisions with.
Defaults to None, in which case no global prim paths are considered.
"""
# validate paths in global prim paths
if global_prim_paths is None:
global_prim_paths = []
else:
# remove duplicates in paths
global_prim_paths = list(set(global_prim_paths))
# set global prim paths list if not previously defined
if len(self._global_prim_paths) < 1:
self._global_prim_paths += global_prim_paths
# filter collisions within each environment instance
self.cloner.filter_collisions(
self.physics_scene_path,
"/World/collisions",
self.env_prim_paths,
global_paths=self._global_prim_paths,
)
def __str__(self) -> str:
"""Returns a string representation of the scene."""
msg = f"<class {self.__class__.__name__}>\n"
msg += f"\tNumber of environments: {self.cfg.num_envs}\n"
msg += f"\tEnvironment spacing : {self.cfg.env_spacing}\n"
msg += f"\tSource prim name : {self.env_prim_paths[0]}\n"
msg += f"\tGlobal prim paths : {self._global_prim_paths}\n"
msg += f"\tReplicate physics : {self.cfg.replicate_physics}"
return msg
"""
Properties.
"""
@property
def physics_scene_path(self) -> str:
"""The path to the USD Physics Scene."""
if self._physics_scene_path is None:
for prim in self.stage.Traverse():
if prim.HasAPI(PhysxSchema.PhysxSceneAPI):
self._physics_scene_path = prim.GetPrimPath().pathString
omni.log.info(f"Physics scene prim path: {self._physics_scene_path}")
break
if self._physics_scene_path is None:
raise RuntimeError("No physics scene found! Please make sure one exists.")
return self._physics_scene_path
@property
def physics_dt(self) -> float:
"""The physics timestep of the scene."""
return sim_utils.SimulationContext.instance().get_physics_dt() # pyright: ignore [reportOptionalMemberAccess]
@property
def device(self) -> str:
"""The device on which the scene is created."""
return sim_utils.SimulationContext.instance().device # pyright: ignore [reportOptionalMemberAccess]
@property
def env_ns(self) -> str:
"""The namespace ``/World/envs`` in which all environments created.
The environments are present w.r.t. this namespace under "env_{N}" prim,
where N is a natural number.
"""
return "/World/envs"
@property
def env_regex_ns(self) -> str:
"""The namespace ``/World/envs/env_.*`` in which all environments created."""
return f"{self.env_ns}/env_.*"
@property
def num_envs(self) -> int:
"""The number of environments handled by the scene."""
return self.cfg.num_envs
@property
def env_origins(self) -> torch.Tensor:
"""The origins of the environments in the scene. Shape is (num_envs, 3)."""
if self._terrain is not None:
return self._terrain.env_origins
else:
return self._default_env_origins
@property
def terrain(self) -> TerrainImporter | None:
"""The terrain in the scene. If None, then the scene has no terrain.
Note:
We treat terrain separate from :attr:`extras` since terrains define environment origins and are
handled differently from other miscellaneous entities.
"""
return self._terrain
@property
def articulations(self) -> dict[str, Articulation]:
"""A dictionary of articulations in the scene."""
return self._articulations
@property
def deformable_objects(self) -> dict[str, DeformableObject]:
"""A dictionary of deformable objects in the scene."""
return self._deformable_objects
@property
def rigid_objects(self) -> dict[str, RigidObject]:
"""A dictionary of rigid objects in the scene."""
return self._rigid_objects
@property
def rigid_object_collections(self) -> dict[str, RigidObjectCollection]:
"""A dictionary of rigid object collections in the scene."""
return self._rigid_object_collections
@property
def sensors(self) -> dict[str, SensorBase]:
"""A dictionary of the sensors in the scene, such as cameras and contact reporters."""
return self._sensors
@property
def extras(self) -> dict[str, XFormPrim]:
"""A dictionary of miscellaneous simulation objects that neither inherit from assets nor sensors.
The keys are the names of the miscellaneous objects, and the values are the `XFormPrim`_
of the corresponding prims.
As an example, lights or other props in the scene that do not have any attributes or properties that you
want to alter at runtime can be added to this dictionary.
Note:
These are not reset or updated by the scene. They are mainly other prims that are not necessarily
handled by the interactive scene, but are useful to be accessed by the user.
.. _XFormPrim: https://docs.omniverse.nvidia.com/py/isaacsim/source/isaacsim.core/docs/index.html#isaacsim.core.prims.XFormPrim
"""
return self._extras
@property
def state(self) -> dict[str, dict[str, dict[str, torch.Tensor]]]:
"""A dictionary of the state of the scene entities in the simulation world frame.
Please refer to :meth:`get_state` for the format.
"""
return self.get_state(is_relative=False)
"""
Operations.
"""
[docs] def reset(self, env_ids: Sequence[int] | None = None):
"""Resets the scene entities.
Args:
env_ids: The indices of the environments to reset.
Defaults to None (all instances).
"""
# -- assets
for articulation in self._articulations.values():
articulation.reset(env_ids)
for deformable_object in self._deformable_objects.values():
deformable_object.reset(env_ids)
for rigid_object in self._rigid_objects.values():
rigid_object.reset(env_ids)
for rigid_object_collection in self._rigid_object_collections.values():
rigid_object_collection.reset(env_ids)
# -- sensors
for sensor in self._sensors.values():
sensor.reset(env_ids)
[docs] def write_data_to_sim(self):
"""Writes the data of the scene entities to the simulation."""
# -- assets
for articulation in self._articulations.values():
articulation.write_data_to_sim()
for deformable_object in self._deformable_objects.values():
deformable_object.write_data_to_sim()
for rigid_object in self._rigid_objects.values():
rigid_object.write_data_to_sim()
for rigid_object_collection in self._rigid_object_collections.values():
rigid_object_collection.write_data_to_sim()
[docs] def update(self, dt: float) -> None:
"""Update the scene entities.
Args:
dt: The amount of time passed from last :meth:`update` call.
"""
# -- assets
for articulation in self._articulations.values():
articulation.update(dt)
for deformable_object in self._deformable_objects.values():
deformable_object.update(dt)
for rigid_object in self._rigid_objects.values():
rigid_object.update(dt)
for rigid_object_collection in self._rigid_object_collections.values():
rigid_object_collection.update(dt)
# -- sensors
for sensor in self._sensors.values():
sensor.update(dt, force_recompute=not self.cfg.lazy_sensor_update)
"""
Operations: Scene State.
"""
[docs] def reset_to(
self,
state: dict[str, dict[str, dict[str, torch.Tensor]]],
env_ids: Sequence[int] | None = None,
is_relative: bool = False,
):
"""Resets the entities in the scene to the provided state.
Args:
state: The state to reset the scene entities to. Please refer to :meth:`get_state` for the format.
env_ids: The indices of the environments to reset. Defaults to None, in which case
all environment instances are reset.
is_relative: If set to True, the state is considered relative to the environment origins.
Defaults to False.
"""
# resolve env_ids
if env_ids is None:
env_ids = slice(None)
# articulations
for asset_name, articulation in self._articulations.items():
asset_state = state["articulation"][asset_name]
# root state
root_pose = asset_state["root_pose"].clone()
if is_relative:
root_pose[:, :3] += self.env_origins[env_ids]
root_velocity = asset_state["root_velocity"].clone()
articulation.write_root_pose_to_sim(root_pose, env_ids=env_ids)
articulation.write_root_velocity_to_sim(root_velocity, env_ids=env_ids)
# joint state
joint_position = asset_state["joint_position"].clone()
joint_velocity = asset_state["joint_velocity"].clone()
articulation.write_joint_state_to_sim(joint_position, joint_velocity, env_ids=env_ids)
# FIXME: This is not generic as it assumes PD control over the joints.
# This assumption does not hold for effort controlled joints.
articulation.set_joint_position_target(joint_position, env_ids=env_ids)
articulation.set_joint_velocity_target(joint_velocity, env_ids=env_ids)
# deformable objects
for asset_name, deformable_object in self._deformable_objects.items():
asset_state = state["deformable_object"][asset_name]
nodal_position = asset_state["nodal_position"].clone()
if is_relative:
nodal_position[:, :3] += self.env_origins[env_ids]
nodal_velocity = asset_state["nodal_velocity"].clone()
deformable_object.write_nodal_pos_to_sim(nodal_position, env_ids=env_ids)
deformable_object.write_nodal_velocity_to_sim(nodal_velocity, env_ids=env_ids)
# rigid objects
for asset_name, rigid_object in self._rigid_objects.items():
asset_state = state["rigid_object"][asset_name]
root_pose = asset_state["root_pose"].clone()
if is_relative:
root_pose[:, :3] += self.env_origins[env_ids]
root_velocity = asset_state["root_velocity"].clone()
rigid_object.write_root_pose_to_sim(root_pose, env_ids=env_ids)
rigid_object.write_root_velocity_to_sim(root_velocity, env_ids=env_ids)
# write data to simulation to make sure initial state is set
# this propagates the joint targets to the simulation
self.write_data_to_sim()
[docs] def get_state(self, is_relative: bool = False) -> dict[str, dict[str, dict[str, torch.Tensor]]]:
"""Returns the state of the scene entities.
Based on the type of the entity, the state comprises of different components.
* For an articulation, the state comprises of the root pose, root velocity, and joint position and velocity.
* For a deformable object, the state comprises of the nodal position and velocity.
* For a rigid object, the state comprises of the root pose and root velocity.
The returned state is a dictionary with the following format:
.. code-block:: python
{
"articulation": {
"entity_1_name": {
"root_pose": torch.Tensor,
"root_velocity": torch.Tensor,
"joint_position": torch.Tensor,
"joint_velocity": torch.Tensor,
},
"entity_2_name": {
"root_pose": torch.Tensor,
"root_velocity": torch.Tensor,
"joint_position": torch.Tensor,
"joint_velocity": torch.Tensor,
},
},
"deformable_object": {
"entity_3_name": {
"nodal_position": torch.Tensor,
"nodal_velocity": torch.Tensor,
}
},
"rigid_object": {
"entity_4_name": {
"root_pose": torch.Tensor,
"root_velocity": torch.Tensor,
}
},
}
where ``entity_N_name`` is the name of the entity registered in the scene.
Args:
is_relative: If set to True, the state is considered relative to the environment origins.
Defaults to False.
Returns:
A dictionary of the state of the scene entities.
"""
state = dict()
# articulations
state["articulation"] = dict()
for asset_name, articulation in self._articulations.items():
asset_state = dict()
asset_state["root_pose"] = articulation.data.root_state_w[:, :7].clone()
if is_relative:
asset_state["root_pose"][:, :3] -= self.env_origins
asset_state["root_velocity"] = articulation.data.root_vel_w.clone()
asset_state["joint_position"] = articulation.data.joint_pos.clone()
asset_state["joint_velocity"] = articulation.data.joint_vel.clone()
state["articulation"][asset_name] = asset_state
# deformable objects
state["deformable_object"] = dict()
for asset_name, deformable_object in self._deformable_objects.items():
asset_state = dict()
asset_state["nodal_position"] = deformable_object.data.nodal_pos_w.clone()
if is_relative:
asset_state["nodal_position"][:, :3] -= self.env_origins
asset_state["nodal_velocity"] = deformable_object.data.nodal_vel_w.clone()
state["deformable_object"][asset_name] = asset_state
# rigid objects
state["rigid_object"] = dict()
for asset_name, rigid_object in self._rigid_objects.items():
asset_state = dict()
asset_state["root_pose"] = rigid_object.data.root_state_w[:, :7].clone()
if is_relative:
asset_state["root_pose"][:, :3] -= self.env_origins
asset_state["root_velocity"] = rigid_object.data.root_vel_w.clone()
state["rigid_object"][asset_name] = asset_state
return state
"""
Operations: Iteration.
"""
[docs] def keys(self) -> list[str]:
"""Returns the keys of the scene entities.
Returns:
The keys of the scene entities.
"""
all_keys = ["terrain"]
for asset_family in [
self._articulations,
self._deformable_objects,
self._rigid_objects,
self._rigid_object_collections,
self._sensors,
self._extras,
]:
all_keys += list(asset_family.keys())
return all_keys
def __getitem__(self, key: str) -> Any:
"""Returns the scene entity with the given key.
Args:
key: The key of the scene entity.
Returns:
The scene entity.
"""
# check if it is a terrain
if key == "terrain":
return self._terrain
all_keys = ["terrain"]
# check if it is in other dictionaries
for asset_family in [
self._articulations,
self._deformable_objects,
self._rigid_objects,
self._rigid_object_collections,
self._sensors,
self._extras,
]:
out = asset_family.get(key)
# if found, return
if out is not None:
return out
all_keys += list(asset_family.keys())
# if not found, raise error
raise KeyError(f"Scene entity with key '{key}' not found. Available Entities: '{all_keys}'")
"""
Internal methods.
"""
def _is_scene_setup_from_cfg(self):
return any(
not (asset_name in InteractiveSceneCfg.__dataclass_fields__ or asset_cfg is None)
for asset_name, asset_cfg in self.cfg.__dict__.items()
)
def _add_entities_from_cfg(self):
"""Add scene entities from the config."""
# store paths that are in global collision filter
self._global_prim_paths = list()
# parse the entire scene config and resolve regex
for asset_name, asset_cfg in self.cfg.__dict__.items():
# skip keywords
# note: easier than writing a list of keywords: [num_envs, env_spacing, lazy_sensor_update]
if asset_name in InteractiveSceneCfg.__dataclass_fields__ or asset_cfg is None:
continue
# resolve regex
if hasattr(asset_cfg, "prim_path"):
asset_cfg.prim_path = asset_cfg.prim_path.format(ENV_REGEX_NS=self.env_regex_ns)
# create asset
if isinstance(asset_cfg, TerrainImporterCfg):
# terrains are special entities since they define environment origins
asset_cfg.num_envs = self.cfg.num_envs
asset_cfg.env_spacing = self.cfg.env_spacing
self._terrain = asset_cfg.class_type(asset_cfg)
elif isinstance(asset_cfg, ArticulationCfg):
self._articulations[asset_name] = asset_cfg.class_type(asset_cfg)
elif isinstance(asset_cfg, DeformableObjectCfg):
self._deformable_objects[asset_name] = asset_cfg.class_type(asset_cfg)
elif isinstance(asset_cfg, RigidObjectCfg):
self._rigid_objects[asset_name] = asset_cfg.class_type(asset_cfg)
elif isinstance(asset_cfg, RigidObjectCollectionCfg):
for rigid_object_cfg in asset_cfg.rigid_objects.values():
rigid_object_cfg.prim_path = rigid_object_cfg.prim_path.format(ENV_REGEX_NS=self.env_regex_ns)
self._rigid_object_collections[asset_name] = asset_cfg.class_type(asset_cfg)
for rigid_object_cfg in asset_cfg.rigid_objects.values():
if hasattr(rigid_object_cfg, "collision_group") and rigid_object_cfg.collision_group == -1:
asset_paths = sim_utils.find_matching_prim_paths(rigid_object_cfg.prim_path)
self._global_prim_paths += asset_paths
elif isinstance(asset_cfg, SensorBaseCfg):
# Update target frame path(s)' regex name space for FrameTransformer
if isinstance(asset_cfg, FrameTransformerCfg):
updated_target_frames = []
for target_frame in asset_cfg.target_frames:
target_frame.prim_path = target_frame.prim_path.format(ENV_REGEX_NS=self.env_regex_ns)
updated_target_frames.append(target_frame)
asset_cfg.target_frames = updated_target_frames
elif isinstance(asset_cfg, ContactSensorCfg):
updated_filter_prim_paths_expr = []
for filter_prim_path in asset_cfg.filter_prim_paths_expr:
updated_filter_prim_paths_expr.append(filter_prim_path.format(ENV_REGEX_NS=self.env_regex_ns))
asset_cfg.filter_prim_paths_expr = updated_filter_prim_paths_expr
self._sensors[asset_name] = asset_cfg.class_type(asset_cfg)
elif isinstance(asset_cfg, AssetBaseCfg):
# manually spawn asset
if asset_cfg.spawn is not None:
asset_cfg.spawn.func(
asset_cfg.prim_path,
asset_cfg.spawn,
translation=asset_cfg.init_state.pos,
orientation=asset_cfg.init_state.rot,
)
# store xform prim view corresponding to this asset
# all prims in the scene are Xform prims (i.e. have a transform component)
self._extras[asset_name] = XFormPrim(asset_cfg.prim_path, reset_xform_properties=False)
else:
raise ValueError(f"Unknown asset config type for {asset_name}: {asset_cfg}")
# store global collision paths
if hasattr(asset_cfg, "collision_group") and asset_cfg.collision_group == -1:
asset_paths = sim_utils.find_matching_prim_paths(asset_cfg.prim_path)
self._global_prim_paths += asset_paths
