# Copyright (c) 2022-2025, The Isaac Lab Project Developers.
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause
from __future__ import annotations
import builtins
import gymnasium as gym
import inspect
import math
import numpy as np
import torch
import weakref
from abc import abstractmethod
from collections.abc import Sequence
from dataclasses import MISSING
from typing import Any, ClassVar
import isaacsim.core.utils.torch as torch_utils
import omni.kit.app
import omni.log
from isaacsim.core.simulation_manager import SimulationManager
from isaacsim.core.version import get_version
from isaaclab.managers import EventManager
from isaaclab.scene import InteractiveScene
from isaaclab.sim import SimulationContext
from isaaclab.utils.noise import NoiseModel
from isaaclab.utils.timer import Timer
from .common import VecEnvObs, VecEnvStepReturn
from .direct_rl_env_cfg import DirectRLEnvCfg
from .ui import ViewportCameraController
from .utils.spaces import sample_space, spec_to_gym_space
[docs]class DirectRLEnv(gym.Env):
"""The superclass for the direct workflow to design environments.
This class implements the core functionality for reinforcement learning (RL)
environments. It is designed to be used with any RL library. The class is designed
to be used with vectorized environments, i.e., the environment is expected to be run
in parallel with multiple sub-environments.
While the environment itself is implemented as a vectorized environment, we do not
inherit from :class:`gym.vector.VectorEnv`. This is mainly because the class adds
various methods (for wait and asynchronous updates) which are not required.
Additionally, each RL library typically has its own definition for a vectorized
environment. Thus, to reduce complexity, we directly use the :class:`gym.Env` over
here and leave it up to library-defined wrappers to take care of wrapping this
environment for their agents.
Note:
For vectorized environments, it is recommended to **only** call the :meth:`reset`
method once before the first call to :meth:`step`, i.e. after the environment is created.
After that, the :meth:`step` function handles the reset of terminated sub-environments.
This is because the simulator does not support resetting individual sub-environments
in a vectorized environment.
"""
is_vector_env: ClassVar[bool] = True
"""Whether the environment is a vectorized environment."""
metadata: ClassVar[dict[str, Any]] = {
"render_modes": [None, "human", "rgb_array"],
"isaac_sim_version": get_version(),
}
"""Metadata for the environment."""
[docs] def __init__(self, cfg: DirectRLEnvCfg, render_mode: str | None = None, **kwargs):
"""Initialize the environment.
Args:
cfg: The configuration object for the environment.
render_mode: The render mode for the environment. Defaults to None, which
is similar to ``"human"``.
Raises:
RuntimeError: If a simulation context already exists. The environment must always create one
since it configures the simulation context and controls the simulation.
"""
# check that the config is valid
cfg.validate()
# store inputs to class
self.cfg = cfg
# store the render mode
self.render_mode = render_mode
# initialize internal variables
self._is_closed = False
# set the seed for the environment
if self.cfg.seed is not None:
self.cfg.seed = self.seed(self.cfg.seed)
else:
omni.log.warn("Seed not set for the environment. The environment creation may not be deterministic.")
# create a simulation context to control the simulator
if SimulationContext.instance() is None:
self.sim: SimulationContext = SimulationContext(self.cfg.sim)
else:
raise RuntimeError("Simulation context already exists. Cannot create a new one.")
# make sure torch is running on the correct device
if "cuda" in self.device:
torch.cuda.set_device(self.device)
# print useful information
print("[INFO]: Base environment:")
print(f"\tEnvironment device : {self.device}")
print(f"\tEnvironment seed : {self.cfg.seed}")
print(f"\tPhysics step-size : {self.physics_dt}")
print(f"\tRendering step-size : {self.physics_dt * self.cfg.sim.render_interval}")
print(f"\tEnvironment step-size : {self.step_dt}")
if self.cfg.sim.render_interval < self.cfg.decimation:
msg = (
f"The render interval ({self.cfg.sim.render_interval}) is smaller than the decimation "
f"({self.cfg.decimation}). Multiple render calls will happen for each environment step."
"If this is not intended, set the render interval to be equal to the decimation."
)
omni.log.warn(msg)
# generate scene
with Timer("[INFO]: Time taken for scene creation", "scene_creation"):
self.scene = InteractiveScene(self.cfg.scene)
self._setup_scene()
print("[INFO]: Scene manager: ", self.scene)
# set up camera viewport controller
# viewport is not available in other rendering modes so the function will throw a warning
# FIXME: This needs to be fixed in the future when we unify the UI functionalities even for
# non-rendering modes.
if self.sim.render_mode >= self.sim.RenderMode.PARTIAL_RENDERING:
self.viewport_camera_controller = ViewportCameraController(self, self.cfg.viewer)
else:
self.viewport_camera_controller = None
# create event manager
# note: this is needed here (rather than after simulation play) to allow USD-related randomization events
# that must happen before the simulation starts. Example: randomizing mesh scale
if self.cfg.events:
self.event_manager = EventManager(self.cfg.events, self)
# apply USD-related randomization events
if "prestartup" in self.event_manager.available_modes:
self.event_manager.apply(mode="prestartup")
# play the simulator to activate physics handles
# note: this activates the physics simulation view that exposes TensorAPIs
# note: when started in extension mode, first call sim.reset_async() and then initialize the managers
if builtins.ISAAC_LAUNCHED_FROM_TERMINAL is False:
print("[INFO]: Starting the simulation. This may take a few seconds. Please wait...")
with Timer("[INFO]: Time taken for simulation start", "simulation_start"):
self.sim.reset()
# update scene to pre populate data buffers for assets and sensors.
# this is needed for the observation manager to get valid tensors for initialization.
# this shouldn't cause an issue since later on, users do a reset over all the environments so the lazy buffers would be reset.
self.scene.update(dt=self.physics_dt)
# check if debug visualization is has been implemented by the environment
source_code = inspect.getsource(self._set_debug_vis_impl)
self.has_debug_vis_implementation = "NotImplementedError" not in source_code
self._debug_vis_handle = None
# extend UI elements
# we need to do this here after all the managers are initialized
# this is because they dictate the sensors and commands right now
if self.sim.has_gui() and self.cfg.ui_window_class_type is not None:
self._window = self.cfg.ui_window_class_type(self, window_name="IsaacLab")
else:
# if no window, then we don't need to store the window
self._window = None
# allocate dictionary to store metrics
self.extras = {}
# initialize data and constants
# -- counter for simulation steps
self._sim_step_counter = 0
# -- counter for curriculum
self.common_step_counter = 0
# -- init buffers
self.episode_length_buf = torch.zeros(self.num_envs, device=self.device, dtype=torch.long)
self.reset_terminated = torch.zeros(self.num_envs, device=self.device, dtype=torch.bool)
self.reset_time_outs = torch.zeros_like(self.reset_terminated)
self.reset_buf = torch.zeros(self.num_envs, dtype=torch.bool, device=self.sim.device)
# setup the action and observation spaces for Gym
self._configure_gym_env_spaces()
# setup noise cfg for adding action and observation noise
if self.cfg.action_noise_model:
self._action_noise_model: NoiseModel = self.cfg.action_noise_model.class_type(
self.cfg.action_noise_model, num_envs=self.num_envs, device=self.device
)
if self.cfg.observation_noise_model:
self._observation_noise_model: NoiseModel = self.cfg.observation_noise_model.class_type(
self.cfg.observation_noise_model, num_envs=self.num_envs, device=self.device
)
# perform events at the start of the simulation
if self.cfg.events:
# we print it here to make the logging consistent
print("[INFO] Event Manager: ", self.event_manager)
if "startup" in self.event_manager.available_modes:
self.event_manager.apply(mode="startup")
# set the framerate of the gym video recorder wrapper so that the playback speed of the produced
# video matches the simulation
self.metadata["render_fps"] = 1 / self.step_dt
# print the environment information
print("[INFO]: Completed setting up the environment...")
def __del__(self):
"""Cleanup for the environment."""
self.close()
"""
Properties.
"""
@property
def num_envs(self) -> int:
"""The number of instances of the environment that are running."""
return self.scene.num_envs
@property
def physics_dt(self) -> float:
"""The physics time-step (in s).
This is the lowest time-decimation at which the simulation is happening.
"""
return self.cfg.sim.dt
@property
def step_dt(self) -> float:
"""The environment stepping time-step (in s).
This is the time-step at which the environment steps forward.
"""
return self.cfg.sim.dt * self.cfg.decimation
@property
def device(self):
"""The device on which the environment is running."""
return self.sim.device
@property
def max_episode_length_s(self) -> float:
"""Maximum episode length in seconds."""
return self.cfg.episode_length_s
@property
def max_episode_length(self):
"""The maximum episode length in steps adjusted from s."""
return math.ceil(self.max_episode_length_s / (self.cfg.sim.dt * self.cfg.decimation))
"""
Operations.
"""
[docs] def reset(self, seed: int | None = None, options: dict[str, Any] | None = None) -> tuple[VecEnvObs, dict]:
"""Resets all the environments and returns observations.
This function calls the :meth:`_reset_idx` function to reset all the environments.
However, certain operations, such as procedural terrain generation, that happened during initialization
are not repeated.
Args:
seed: The seed to use for randomization. Defaults to None, in which case the seed is not set.
options: Additional information to specify how the environment is reset. Defaults to None.
Note:
This argument is used for compatibility with Gymnasium environment definition.
Returns:
A tuple containing the observations and extras.
"""
# set the seed
if seed is not None:
self.seed(seed)
# reset state of scene
indices = torch.arange(self.num_envs, dtype=torch.int64, device=self.device)
self._reset_idx(indices)
# update articulation kinematics
self.scene.write_data_to_sim()
self.sim.forward()
# if sensors are added to the scene, make sure we render to reflect changes in reset
if self.sim.has_rtx_sensors() and self.cfg.rerender_on_reset:
self.sim.render()
if self.cfg.wait_for_textures and self.sim.has_rtx_sensors():
while SimulationManager.assets_loading():
self.sim.render()
# return observations
return self._get_observations(), self.extras
[docs] def step(self, action: torch.Tensor) -> VecEnvStepReturn:
"""Execute one time-step of the environment's dynamics.
The environment steps forward at a fixed time-step, while the physics simulation is decimated at a
lower time-step. This is to ensure that the simulation is stable. These two time-steps can be configured
independently using the :attr:`DirectRLEnvCfg.decimation` (number of simulation steps per environment step)
and the :attr:`DirectRLEnvCfg.sim.physics_dt` (physics time-step). Based on these parameters, the environment
time-step is computed as the product of the two.
This function performs the following steps:
1. Pre-process the actions before stepping through the physics.
2. Apply the actions to the simulator and step through the physics in a decimated manner.
3. Compute the reward and done signals.
4. Reset environments that have terminated or reached the maximum episode length.
5. Apply interval events if they are enabled.
6. Compute observations.
Args:
action: The actions to apply on the environment. Shape is (num_envs, action_dim).
Returns:
A tuple containing the observations, rewards, resets (terminated and truncated) and extras.
"""
action = action.to(self.device)
# add action noise
if self.cfg.action_noise_model:
action = self._action_noise_model.apply(action)
# process actions
self._pre_physics_step(action)
# check if we need to do rendering within the physics loop
# note: checked here once to avoid multiple checks within the loop
is_rendering = self.sim.has_gui() or self.sim.has_rtx_sensors()
# perform physics stepping
for _ in range(self.cfg.decimation):
self._sim_step_counter += 1
# set actions into buffers
self._apply_action()
# set actions into simulator
self.scene.write_data_to_sim()
# simulate
self.sim.step(render=False)
# render between steps only if the GUI or an RTX sensor needs it
# note: we assume the render interval to be the shortest accepted rendering interval.
# If a camera needs rendering at a faster frequency, this will lead to unexpected behavior.
if self._sim_step_counter % self.cfg.sim.render_interval == 0 and is_rendering:
self.sim.render()
# update buffers at sim dt
self.scene.update(dt=self.physics_dt)
# post-step:
# -- update env counters (used for curriculum generation)
self.episode_length_buf += 1 # step in current episode (per env)
self.common_step_counter += 1 # total step (common for all envs)
self.reset_terminated[:], self.reset_time_outs[:] = self._get_dones()
self.reset_buf = self.reset_terminated | self.reset_time_outs
self.reward_buf = self._get_rewards()
# -- reset envs that terminated/timed-out and log the episode information
reset_env_ids = self.reset_buf.nonzero(as_tuple=False).squeeze(-1)
if len(reset_env_ids) > 0:
self._reset_idx(reset_env_ids)
# update articulation kinematics
self.scene.write_data_to_sim()
self.sim.forward()
# if sensors are added to the scene, make sure we render to reflect changes in reset
if self.sim.has_rtx_sensors() and self.cfg.rerender_on_reset:
self.sim.render()
# post-step: step interval event
if self.cfg.events:
if "interval" in self.event_manager.available_modes:
self.event_manager.apply(mode="interval", dt=self.step_dt)
# update observations
self.obs_buf = self._get_observations()
# add observation noise
# note: we apply no noise to the state space (since it is used for critic networks)
if self.cfg.observation_noise_model:
self.obs_buf["policy"] = self._observation_noise_model.apply(self.obs_buf["policy"])
# return observations, rewards, resets and extras
return self.obs_buf, self.reward_buf, self.reset_terminated, self.reset_time_outs, self.extras
[docs] @staticmethod
def seed(seed: int = -1) -> int:
"""Set the seed for the environment.
Args:
seed: The seed for random generator. Defaults to -1.
Returns:
The seed used for random generator.
"""
# set seed for replicator
try:
import omni.replicator.core as rep
rep.set_global_seed(seed)
except ModuleNotFoundError:
pass
# set seed for torch and other libraries
return torch_utils.set_seed(seed)
[docs] def render(self, recompute: bool = False) -> np.ndarray | None:
"""Run rendering without stepping through the physics.
By convention, if mode is:
- **human**: Render to the current display and return nothing. Usually for human consumption.
- **rgb_array**: Return an numpy.ndarray with shape (x, y, 3), representing RGB values for an
x-by-y pixel image, suitable for turning into a video.
Args:
recompute: Whether to force a render even if the simulator has already rendered the scene.
Defaults to False.
Returns:
The rendered image as a numpy array if mode is "rgb_array". Otherwise, returns None.
Raises:
RuntimeError: If mode is set to "rgb_data" and simulation render mode does not support it.
In this case, the simulation render mode must be set to ``RenderMode.PARTIAL_RENDERING``
or ``RenderMode.FULL_RENDERING``.
NotImplementedError: If an unsupported rendering mode is specified.
"""
# run a rendering step of the simulator
# if we have rtx sensors, we do not need to render again sin
if not self.sim.has_rtx_sensors() and not recompute:
self.sim.render()
# decide the rendering mode
if self.render_mode == "human" or self.render_mode is None:
return None
elif self.render_mode == "rgb_array":
# check that if any render could have happened
if self.sim.render_mode.value < self.sim.RenderMode.PARTIAL_RENDERING.value:
raise RuntimeError(
f"Cannot render '{self.render_mode}' when the simulation render mode is"
f" '{self.sim.render_mode.name}'. Please set the simulation render mode to:"
f"'{self.sim.RenderMode.PARTIAL_RENDERING.name}' or '{self.sim.RenderMode.FULL_RENDERING.name}'."
" If running headless, make sure --enable_cameras is set."
)
# create the annotator if it does not exist
if not hasattr(self, "_rgb_annotator"):
import omni.replicator.core as rep
# create render product
self._render_product = rep.create.render_product(
self.cfg.viewer.cam_prim_path, self.cfg.viewer.resolution
)
# create rgb annotator -- used to read data from the render product
self._rgb_annotator = rep.AnnotatorRegistry.get_annotator("rgb", device="cpu")
self._rgb_annotator.attach([self._render_product])
# obtain the rgb data
rgb_data = self._rgb_annotator.get_data()
# convert to numpy array
rgb_data = np.frombuffer(rgb_data, dtype=np.uint8).reshape(*rgb_data.shape)
# return the rgb data
# note: initially the renerer is warming up and returns empty data
if rgb_data.size == 0:
return np.zeros((self.cfg.viewer.resolution[1], self.cfg.viewer.resolution[0], 3), dtype=np.uint8)
else:
return rgb_data[:, :, :3]
else:
raise NotImplementedError(
f"Render mode '{self.render_mode}' is not supported. Please use: {self.metadata['render_modes']}."
)
[docs] def close(self):
"""Cleanup for the environment."""
if not self._is_closed:
# close entities related to the environment
# note: this is order-sensitive to avoid any dangling references
if self.cfg.events:
del self.event_manager
del self.scene
if self.viewport_camera_controller is not None:
del self.viewport_camera_controller
# clear callbacks and instance
self.sim.clear_all_callbacks()
self.sim.clear_instance()
# destroy the window
if self._window is not None:
self._window = None
# update closing status
self._is_closed = True
"""
Operations - Debug Visualization.
"""
[docs] def set_debug_vis(self, debug_vis: bool) -> bool:
"""Toggles the environment debug visualization.
Args:
debug_vis: Whether to visualize the environment debug visualization.
Returns:
Whether the debug visualization was successfully set. False if the environment
does not support debug visualization.
"""
# check if debug visualization is supported
if not self.has_debug_vis_implementation:
return False
# toggle debug visualization objects
self._set_debug_vis_impl(debug_vis)
# toggle debug visualization handles
if debug_vis:
# create a subscriber for the post update event if it doesn't exist
if self._debug_vis_handle is None:
app_interface = omni.kit.app.get_app_interface()
self._debug_vis_handle = app_interface.get_post_update_event_stream().create_subscription_to_pop(
lambda event, obj=weakref.proxy(self): obj._debug_vis_callback(event)
)
else:
# remove the subscriber if it exists
if self._debug_vis_handle is not None:
self._debug_vis_handle.unsubscribe()
self._debug_vis_handle = None
# return success
return True
"""
Helper functions.
"""
def _configure_gym_env_spaces(self):
"""Configure the action and observation spaces for the Gym environment."""
# show deprecation message and overwrite configuration
if self.cfg.num_actions is not None:
omni.log.warn("DirectRLEnvCfg.num_actions is deprecated. Use DirectRLEnvCfg.action_space instead.")
if isinstance(self.cfg.action_space, type(MISSING)):
self.cfg.action_space = self.cfg.num_actions
if self.cfg.num_observations is not None:
omni.log.warn(
"DirectRLEnvCfg.num_observations is deprecated. Use DirectRLEnvCfg.observation_space instead."
)
if isinstance(self.cfg.observation_space, type(MISSING)):
self.cfg.observation_space = self.cfg.num_observations
if self.cfg.num_states is not None:
omni.log.warn("DirectRLEnvCfg.num_states is deprecated. Use DirectRLEnvCfg.state_space instead.")
if isinstance(self.cfg.state_space, type(MISSING)):
self.cfg.state_space = self.cfg.num_states
# set up spaces
self.single_observation_space = gym.spaces.Dict()
self.single_observation_space["policy"] = spec_to_gym_space(self.cfg.observation_space)
self.single_action_space = spec_to_gym_space(self.cfg.action_space)
# batch the spaces for vectorized environments
self.observation_space = gym.vector.utils.batch_space(self.single_observation_space["policy"], self.num_envs)
self.action_space = gym.vector.utils.batch_space(self.single_action_space, self.num_envs)
# optional state space for asymmetric actor-critic architectures
self.state_space = None
if self.cfg.state_space:
self.single_observation_space["critic"] = spec_to_gym_space(self.cfg.state_space)
self.state_space = gym.vector.utils.batch_space(self.single_observation_space["critic"], self.num_envs)
# instantiate actions (needed for tasks for which the observations computation is dependent on the actions)
self.actions = sample_space(self.single_action_space, self.sim.device, batch_size=self.num_envs, fill_value=0)
def _reset_idx(self, env_ids: Sequence[int]):
"""Reset environments based on specified indices.
Args:
env_ids: List of environment ids which must be reset
"""
self.scene.reset(env_ids)
# apply events such as randomization for environments that need a reset
if self.cfg.events:
if "reset" in self.event_manager.available_modes:
env_step_count = self._sim_step_counter // self.cfg.decimation
self.event_manager.apply(mode="reset", env_ids=env_ids, global_env_step_count=env_step_count)
# reset noise models
if self.cfg.action_noise_model:
self._action_noise_model.reset(env_ids)
if self.cfg.observation_noise_model:
self._observation_noise_model.reset(env_ids)
# reset the episode length buffer
self.episode_length_buf[env_ids] = 0
"""
Implementation-specific functions.
"""
def _setup_scene(self):
"""Setup the scene for the environment.
This function is responsible for creating the scene objects and setting up the scene for the environment.
The scene creation can happen through :class:`isaaclab.scene.InteractiveSceneCfg` or through
directly creating the scene objects and registering them with the scene manager.
We leave the implementation of this function to the derived classes. If the environment does not require
any explicit scene setup, the function can be left empty.
"""
pass
@abstractmethod
def _pre_physics_step(self, actions: torch.Tensor):
"""Pre-process actions before stepping through the physics.
This function is responsible for pre-processing the actions before stepping through the physics.
It is called before the physics stepping (which is decimated).
Args:
actions: The actions to apply on the environment. Shape is (num_envs, action_dim).
"""
raise NotImplementedError(f"Please implement the '_pre_physics_step' method for {self.__class__.__name__}.")
@abstractmethod
def _apply_action(self):
"""Apply actions to the simulator.
This function is responsible for applying the actions to the simulator. It is called at each
physics time-step.
"""
raise NotImplementedError(f"Please implement the '_apply_action' method for {self.__class__.__name__}.")
@abstractmethod
def _get_observations(self) -> VecEnvObs:
"""Compute and return the observations for the environment.
Returns:
The observations for the environment.
"""
raise NotImplementedError(f"Please implement the '_get_observations' method for {self.__class__.__name__}.")
def _get_states(self) -> VecEnvObs | None:
"""Compute and return the states for the environment.
The state-space is used for asymmetric actor-critic architectures. It is configured
using the :attr:`DirectRLEnvCfg.state_space` parameter.
Returns:
The states for the environment. If the environment does not have a state-space, the function
returns a None.
"""
return None # noqa: R501
@abstractmethod
def _get_rewards(self) -> torch.Tensor:
"""Compute and return the rewards for the environment.
Returns:
The rewards for the environment. Shape is (num_envs,).
"""
raise NotImplementedError(f"Please implement the '_get_rewards' method for {self.__class__.__name__}.")
@abstractmethod
def _get_dones(self) -> tuple[torch.Tensor, torch.Tensor]:
"""Compute and return the done flags for the environment.
Returns:
A tuple containing the done flags for termination and time-out.
Shape of individual tensors is (num_envs,).
"""
raise NotImplementedError(f"Please implement the '_get_dones' method for {self.__class__.__name__}.")
def _set_debug_vis_impl(self, debug_vis: bool):
"""Set debug visualization into visualization objects.
This function is responsible for creating the visualization objects if they don't exist
and input ``debug_vis`` is True. If the visualization objects exist, the function should
set their visibility into the stage.
"""
raise NotImplementedError(f"Debug visualization is not implemented for {self.__class__.__name__}.")
