From IsaacGymEnvs

From IsaacGymEnvs#

IsaacGymEnvs was a reinforcement learning framework designed for the Isaac Gym Preview Release. As both IsaacGymEnvs and the Isaac Gym Preview Release are now deprecated, the following guide walks through the key differences between IsaacGymEnvs and Isaac Lab, as well as differences in APIs between Isaac Gym Preview Release and Isaac Sim.

Note

The following changes are with respect to Isaac Lab 1.0 release. Please refer to the release notes for any changes in the future releases.

Task Config Setup#

In IsaacGymEnvs, task config files were defined in .yaml format. With Isaac Lab, configs are now specified using a specialized Python class configclass. The configclass module provides a wrapper on top of Python’s dataclasses module. Each environment should specify its own config class annotated by @configclass that inherits from DirectRLEnvCfg, which can include simulation parameters, environment scene parameters, robot parameters, and task-specific parameters.

Below is an example skeleton of a task config class:

from omni.isaac.lab.envs import DirectRLEnvCfg
from omni.isaac.lab.scene import InteractiveSceneCfg
from omni.isaac.lab.sim import SimulationCfg

@configclass
class MyEnvCfg(DirectRLEnvCfg):
   # simulation
   sim: SimulationCfg = SimulationCfg()
   # robot
   robot_cfg: ArticulationCfg = ArticulationCfg()
   # scene
   scene: InteractiveSceneCfg = InteractiveSceneCfg()
   # env
   decimation = 2
   episode_length_s = 5.0
   action_space = 1
   observation_space = 4
   state_space = 0
   # task-specific parameters
   ...

Simulation Config#

Simulation related parameters are defined as part of the SimulationCfg class, which is a configclass module that holds simulation parameters such as dt, device, and gravity. Each task config must have a variable named sim defined that holds the type SimulationCfg.

In Isaac Lab, the use of substeps has been replaced by a combination of the simulation dt and the decimation parameters. For example, in IsaacGymEnvs, having dt=1/60 and substeps=2 is equivalent to taking 2 simulation steps with dt=1/120, but running the task step at 1/60 seconds. The decimation parameter is a task parameter that controls the number of simulation steps to take for each task (or RL) step, replacing the controlFrequencyInv parameter in IsaacGymEnvs. Thus, the same setup in Isaac Lab will become dt=1/120 and decimation=2.

In Isaac Sim, physx simulation parameters such as num_position_iterations, num_velocity_iterations, contact_offset, rest_offset, bounce_threshold_velocity, max_depenetration_velocity can all be specified on a per-actor basis. These parameters have been moved from the physx simulation config to each individual articulation and rigid body config.

When running simulation on the GPU, buffers in PhysX require pre-allocation for computing and storing information such as contacts, collisions and aggregate pairs. These buffers may need to be adjusted depending on the complexity of the environment, the number of expected contacts and collisions, and the number of actors in the environment. The PhysxCfg class provides access for setting the GPU buffer dimensions.

# IsaacGymEnvs
sim:

  dt: 0.0166 # 1/60 s
  substeps: 2
  up_axis: "z"
  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
  gravity: [0.0, 0.0, -9.81]
  physx:
    num_threads: ${....num_threads}
    solver_type: ${....solver_type}
    use_gpu: ${contains:"cuda",${....sim_device}}
    num_position_iterations: 4
    num_velocity_iterations: 0
    contact_offset: 0.02
    rest_offset: 0.001
    bounce_threshold_velocity: 0.2
    max_depenetration_velocity: 100.0
    default_buffer_size_multiplier: 2.0
    max_gpu_contact_pairs: 1048576 # 1024*1024
    num_subscenes: ${....num_subscenes}
    contact_collection: 0

# IsaacLab
sim: SimulationCfg = SimulationCfg(
   device = "cuda:0" # can be "cpu", "cuda", "cuda:<device_id>"
   dt=1 / 120,
   # decimation will be set in the task config
   # up axis will always be Z in isaac sim
   # use_gpu_pipeline is deduced from the device
   gravity=(0.0, 0.0, -9.81),
   physx: PhysxCfg = PhysxCfg(
       # num_threads is no longer needed
       solver_type=1,
       # use_gpu is deduced from the device
       max_position_iteration_count=4,
       max_velocity_iteration_count=0,
       # moved to actor config
       # moved to actor config
       bounce_threshold_velocity=0.2,
       # moved to actor config
       # default_buffer_size_multiplier is no longer needed
       gpu_max_rigid_contact_count=2**23
       # num_subscenes is no longer needed
       # contact_collection is no longer needed
))

Scene Config#

The InteractiveSceneCfg class can be used to specify parameters related to the scene, such as the number of environments and the spacing between environments. Each task config must have a variable named scene defined that holds the type InteractiveSceneCfg.

# IsaacGymEnvs
env:
  numEnvs: ${resolve_default:512,${...num_envs}}
  envSpacing: 4.0

# IsaacLab
scene: InteractiveSceneCfg = InteractiveSceneCfg(
   num_envs=512,
   env_spacing=4.0)

Task Config#

Each environment should specify its own config class that holds task specific parameters, such as the dimensions of the observation and action buffers. Reward term scaling parameters can also be specified in the config class.

The following parameters must be set for each environment config:

decimation = 2
episode_length_s = 5.0
action_space = 1
observation_space = 4
state_space = 0

Note that the maximum episode length parameter (now episode_length_s) is in seconds instead of steps as it was in IsaacGymEnvs. To convert between step count to seconds, use the equation: episode_length_s = dt * decimation * num_steps

RL Config Setup#

RL config files for the rl_games library can continue to be defined in .yaml files in Isaac Lab. Most of the content of the config file can be copied directly from IsaacGymEnvs. Note that in Isaac Lab, we do not use hydra to resolve relative paths in config files. Please replace any relative paths such as ${....device} with the actual values of the parameters.

Additionally, the observation and action clip ranges have been moved to the RL config file. For any clipObservations and clipActions parameters that were defined in the IsaacGymEnvs task config file, they should be moved to the RL config file in Isaac Lab.

IsaacGymEnvs Task Config

Isaac Lab RL Config

 
# IsaacGymEnvs
env:
  clipObservations: 5.0
  clipActions: 1.0

# IsaacLab
params:
  env:
    clip_observations: 5.0
    clip_actions: 1.0

Environment Creation#

In IsaacGymEnvs, environment creation generally included four components: creating the sim object with create_sim(), creating the ground plane, importing the assets from MJCF or URDF files, and finally creating the environments by looping through each environment and adding actors into the environments.

Isaac Lab no longer requires calling the create_sim() method to retrieve the sim object. Instead, the simulation context is retrieved automatically by the framework. It is also no longer required to use the sim as an argument for the simulation APIs.

In replacement of create_sim(), tasks can implement the _setup_scene() method in Isaac Lab. This method can be used for adding actors into the scene, adding ground plane, cloning the actors, and adding any other optional objects into the scene, such as lights.

IsaacGymEnvs

Isaac Lab

 
def create_sim(self):
  # set the up axis to be z-up
  self.up_axis = self.cfg["sim"]["up_axis"]

  self.sim = super().create_sim(self.device_id, self.graphics_device_id,
                                  self.physics_engine, self.sim_params)
  self._create_ground_plane()
  self._create_envs(self.num_envs, self.cfg["env"]['envSpacing'],
                      int(np.sqrt(self.num_envs)))

def _setup_scene(self):
  self.cartpole = Articulation(self.cfg.robot_cfg)
  # add ground plane
  spawn_ground_plane(prim_path="/World/ground", cfg=GroundPlaneCfg()
  # clone, filter, and replicate
  self.scene.clone_environments(copy_from_source=False)
  self.scene.filter_collisions(global_prim_paths=[])
  # add articulation to scene
  self.scene.articulations["cartpole"] = self.cartpole
  # add lights
  light_cfg = sim_utils.DomeLightCfg(intensity=2000.0)
  light_cfg.func("/World/Light", light_cfg)

Ground Plane#

In Isaac Lab, most of the environment creation process has been simplified into configs with the configclass module.

The ground plane can be defined using the TerrainImporterCfg class.

from omni.isaac.lab.terrains import TerrainImporterCfg

terrain = TerrainImporterCfg(
     prim_path="/World/ground",
     terrain_type="plane",
     collision_group=-1,
     physics_material=sim_utils.RigidBodyMaterialCfg(
         friction_combine_mode="multiply",
         restitution_combine_mode="multiply",
         static_friction=1.0,
         dynamic_friction=1.0,
         restitution=0.0,
     ),
 )

The terrain can then be added to the scene in _setup_scene(self) by referencing the TerrainImporterCfg object:

Actors#

Isaac Lab and Isaac Sim both use the USD (Universal Scene Description) library for describing the scene. Assets defined in MJCF and URDF formats can be imported to USD using importer tools described in the Importing a New Asset tutorial.

Each Articulation and Rigid Body actor can also have its own config class. The ArticulationCfg class can be used to define parameters for articulation actors, including file path, simulation parameters, actuator properties, and initial states.

Within the ArticulationCfg, the spawn attribute can be used to add the robot to the scene by specifying the path to the robot file. In addition, RigidBodyPropertiesCfg can be used to specify simulation properties for the rigid bodies in the articulation. Similarly, the ArticulationRootPropertiesCfg class can be used to specify simulation properties for the articulation. Joint properties are now specified as part of the actuators dictionary using ImplicitActuatorCfg. Joints with the same properties can be grouped into regex expressions or provided as a list of names or expressions.

Actors are added to the scene by simply calling self.cartpole = Articulation(self.cfg.robot_cfg), where self.cfg.robot_cfg is an ArticulationCfg object. Once initialized, they should also be added to the InteractiveScene by calling self.scene.articulations["cartpole"] = self.cartpole so that the InteractiveScene can traverse through actors in the scene for writing values to the simulation and resetting.

Simulation Parameters for Actors#

Some simulation parameters related to Rigid Bodies and Articulations may have different default values between Isaac Gym Preview Release and Isaac Sim. It may be helpful to double check the USD assets to ensure that the default values are applicable for the asset.

For instance, the following parameters in the RigidBodyAPI could be different between Isaac Gym Preview Release and Isaac Sim:

RigidBodyAPI Parameter

Default Value in Isaac Sim

Default Value in Isaac Gym Preview Release

Linear Damping

0.00

0.00

Angular Damping

0.05

0.0

Max Linear Velocity

inf

1000

Max Angular Velocity

5729.58008 (degree/s)

64.0 (rad/s)

Max Contact Impulse

inf

1e32

Articulation parameters for the JointAPI and DriveAPI could be altered as well. Note that the Isaac Sim UI assumes the unit of angle to be degrees. It is particularly worth noting that the Damping and Stiffness parameters in the DriveAPI have the unit of 1/deg in the Isaac Sim UI but 1/rad in Isaac Gym Preview Release.

Joint Parameter

Default Value in Isaac Sim

Default Value in Isaac Gym Preview Releases

Maximum Joint Velocity

1000000.0 (deg)

100.0 (rad)

Cloner#

Isaac Sim introduced a concept of Cloner, which is a class designed for replication during the scene creation process. In IsaacGymEnvs, scenes had to be created by looping through the number of environments. Within each iteration, actors were added to each environment and their handles had to be cached. Isaac Lab eliminates the need for looping through the environments by using the Cloner APIs. The scene creation process is as follow:

  1. Construct a single environment (what the scene would look like if number of environments = 1)

  2. Call clone_environments() to replicate the single environment

  3. Call filter_collisions() to filter out collision between environments (if required)

# construct a single environment with the Cartpole robot
self.cartpole = Articulation(self.cfg.robot_cfg)
# clone the environment
self.scene.clone_environments(copy_from_source=False)
# filter collisions
self.scene.filter_collisions(global_prim_paths=[self.cfg.terrain.prim_path])

Accessing States from Simulation#

APIs for accessing physics states in Isaac Lab require the creation of an Articulation or RigidObject object. Multiple objects can be initialized for different articulations or rigid bodies in the scene by defining corresponding ArticulationCfg or RigidObjectCfg config as outlined in the section above. This approach eliminates the need of retrieving body handles to slice states for specific bodies in the scene.

self._robot = Articulation(self.cfg.robot)
self._cabinet = Articulation(self.cfg.cabinet)
self._object = RigidObject(self.cfg.object_cfg)

We have also removed acquire and refresh APIs in Isaac Lab. Physics states can be directly applied or retrieved using APIs defined for the articulations and rigid objects.

APIs provided in Isaac Lab no longer require explicit wrapping and un-wrapping of underlying buffers. APIs can now work with tensors directly for reading and writing data.

IsaacGymEnvs

Isaac Lab

 
dof_state_tensor = self.gym.acquire_dof_state_tensor(self.sim)
self.dof_state = gymtorch.wrap_tensor(dof_state_tensor)
self.gym.refresh_dof_state_tensor(self.sim)

self.joint_pos = self._robot.data.joint_pos
self.joint_vel = self._robot.data.joint_vel

Note some naming differences between APIs in Isaac Gym Preview Release and Isaac Lab. Most dof related APIs have been named to joint in Isaac Lab. APIs in Isaac Lab also no longer follow the explicit _tensors or _tensor_indexed suffixes in naming. Indexed versions of APIs now happen implicitly through the optional indices parameter.

Most APIs in Isaac Lab also provide the option to specify an indices parameter, which can be used when reading or writing data for a subset of environments. Note that when setting states with the indices parameter, the shape of the states buffer should match with the dimension of the indices list.

IsaacGymEnvs

Isaac Lab

 
env_ids_int32 = env_ids.to(dtype=torch.int32)
self.gym.set_dof_state_tensor_indexed(self.sim,
    gymtorch.unwrap_tensor(self.dof_state),
    gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))

self._robot.write_joint_state_to_sim(joint_pos, joint_vel,
                                 joint_ids, env_ids)

Quaternion Convention#

Isaac Lab and Isaac Sim both adopt wxyz as the quaternion convention. However, the quaternion convention used in Isaac Gym Preview Release was xyzw. Remember to switch all quaternions to use the xyzw convention when working indexing rotation data. Similarly, please ensure all quaternions are in wxyz before passing them to Isaac Lab APIs.

Articulation Joint Order#

Physics simulation in Isaac Sim and Isaac Lab assumes a breadth-first ordering for the joints in a given kinematic tree. However, Isaac Gym Preview Release assumed a depth-first ordering for joints in the kinematic tree. This means that indexing joints based on their ordering may be different in IsaacGymEnvs and Isaac Lab.

In Isaac Lab, the list of joint names can be retrieved with Articulation.data.joint_names, which will also correspond to the ordering of the joints in the Articulation.

Creating a New Environment#

Each environment in Isaac Lab should be in its own directory following this structure:

my_environment/
    - agents/
        - __init__.py
        - rl_games_ppo_cfg.py
    - __init__.py
    my_env.py

  • my_environment is the root directory of the task.

  • my_environment/agents is the directory containing all RL config files for the task. Isaac Lab supports multiple RL libraries that can each have its own individual config file.

  • my_environment/__init__.py is the main file that registers the environment with the Gymnasium interface. This allows the training and inferencing scripts to find the task by its name. The content of this file should be as follow:

import gymnasium as gym

from . import agents
from .cartpole_env import CartpoleEnv, CartpoleEnvCfg

##
# Register Gym environments.
##

gym.register(
    id="Isaac-Cartpole-Direct-v0",
    entry_point="omni.isaac.lab_tasks.direct_workflow.cartpole:CartpoleEnv",
    disable_env_checker=True,
    kwargs={
        "env_cfg_entry_point": CartpoleEnvCfg,
        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_cfg.yaml"
    },
)

  • my_environment/my_env.py is the main python script that implements the task logic and task config class for the environment.

Task Logic#

In Isaac Lab, the post_physics_step function has been moved to the framework in the base class. Tasks are not required to implement this method, but can choose to override it if a different workflow is desired.

By default, Isaac Lab follows the following flow in logic:

IsaacGymEnvs

Isaac Lab

 
pre_physics_step
  |-- apply_action


post_physics_step
  |-- reset_idx()
  |-- compute_observation()
  |-- compute_reward()

pre_physics_step
  |-- _pre_physics_step(action)
  |-- _apply_action()

post_physics_step
  |-- _get_dones()
  |-- _get_rewards()
  |-- _reset_idx()
  |-- _get_observations()

In Isaac Lab, we also separate the pre_physics_step API for processing actions from the policy with the apply_action API, which sets the actions into the simulation. This provides more flexibility in controlling when actions should be written to simulation when decimation is used. pre_physics_step will be called once per step before stepping simulation. apply_actions will be called decimation number of times for each RL step, once before each simulation step call.

With this approach, resets are performed based on actions from the current step instead of the previous step. Observations will also be computed with the correct states after resets.

We have also performed some renamings of APIs:

  • create_sim(self) –> _setup_scene(self)

  • pre_physics_step(self, actions) –> _pre_physics_step(self, actions) and _apply_action(self)

  • reset_idx(self, env_ids) –> _reset_idx(self, env_ids)

  • compute_observations(self) –> _get_observations(self) - _get_observations() should now return a dictionary {"policy": obs}

  • compute_reward(self) –> _get_rewards(self) - _get_rewards() should now return the reward buffer

  • post_physics_step(self) –> moved to the base class

  • In addition, Isaac Lab requires the implementation of _is_done(self), which should return two buffers: the reset buffer and the time_out buffer.

Putting It All Together#

The Cartpole environment is shown here in completion to fully show the comparison between the IsaacGymEnvs implementation and the Isaac Lab implementation.

Task Config#

IsaacGymEnvs

Isaac Lab

 
# used to create the object
name: Cartpole

physics_engine: ${..physics_engine}

# if given, will override the device setting in gym.
env:
  numEnvs: ${resolve_default:512,${...num_envs}}
  envSpacing: 4.0
  resetDist: 3.0
  maxEffort: 400.0

  clipObservations: 5.0
  clipActions: 1.0

  asset:
    assetRoot: "../../assets"
    assetFileName: "urdf/cartpole.urdf"

  enableCameraSensors: False

sim:
  dt: 0.0166 # 1/60 s
  substeps: 2
  up_axis: "z"
  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
  gravity: [0.0, 0.0, -9.81]
  physx:
    num_threads: ${....num_threads}
    solver_type: ${....solver_type}
    use_gpu: ${contains:"cuda",${....sim_device}}
    num_position_iterations: 4
    num_velocity_iterations: 0
    contact_offset: 0.02
    rest_offset: 0.001
    bounce_threshold_velocity: 0.2
    max_depenetration_velocity: 100.0
    default_buffer_size_multiplier: 2.0
    max_gpu_contact_pairs: 1048576 # 1024*1024
    num_subscenes: ${....num_subscenes}
    contact_collection: 0

@configclass
class CartpoleEnvCfg(DirectRLEnvCfg):

    # simulation
    sim: SimulationCfg = SimulationCfg(dt=1 / 120)
    # robot
    robot_cfg: ArticulationCfg = CARTPOLE_CFG.replace(
        prim_path="/World/envs/env_.*/Robot")
    cart_dof_name = "slider_to_cart"
    pole_dof_name = "cart_to_pole"
    # scene
    scene: InteractiveSceneCfg = InteractiveSceneCfg(
        num_envs=4096, env_spacing=4.0, replicate_physics=True)
    # env
    decimation = 2
    episode_length_s = 5.0
    action_scale = 100.0  # [N]
    action_space = 1
    observation_space = 4
    state_space = 0
    # reset
    max_cart_pos = 3.0
    initial_pole_angle_range = [-0.25, 0.25]
    # reward scales
    rew_scale_alive = 1.0
    rew_scale_terminated = -2.0
    rew_scale_pole_pos = -1.0
    rew_scale_cart_vel = -0.01
    rew_scale_pole_vel = -0.005

Task Setup#

Isaac Lab no longer requires pre-initialization of buffers through the acquire_* APIs that were used in IsaacGymEnvs. It is also no longer necessary to wrap and unwrap tensors.

IsaacGymEnvs

Isaac Lab

 
class Cartpole(VecTask):

  def __init__(self, cfg, rl_device, sim_device, graphics_device_id,
   headless, virtual_screen_capture, force_render):
      self.cfg = cfg

      self.reset_dist = self.cfg["env"]["resetDist"]

      self.max_push_effort = self.cfg["env"]["maxEffort"]
      self.max_episode_length = 500

      self.cfg["env"]["numObservations"] = 4
      self.cfg["env"]["numActions"] = 1

      super().__init__(config=self.cfg,
         rl_device=rl_device, sim_device=sim_device,
         graphics_device_id=graphics_device_id, headless=headless,
         virtual_screen_capture=virtual_screen_capture,
         force_render=force_render)

      dof_state_tensor = self.gym.acquire_dof_state_tensor(self.sim)
      self.dof_state = gymtorch.wrap_tensor(dof_state_tensor)
      self.dof_pos = self.dof_state.view(
          self.num_envs, self.num_dof, 2)[..., 0]
      self.dof_vel = self.dof_state.view(
          self.num_envs, self.num_dof, 2)[..., 1]

class CartpoleEnv(DirectRLEnv):
  cfg: CartpoleEnvCfg
  def __init__(self, cfg: CartpoleEnvCfg,
          render_mode: str | None = None, **kwargs):

      super().__init__(cfg, render_mode, **kwargs)

      self._cart_dof_idx, _ = self.cartpole.find_joints(
          self.cfg.cart_dof_name)
      self._pole_dof_idx, _ = self.cartpole.find_joints(
          self.cfg.pole_dof_name)
      self.action_scale = self.cfg.action_scale

      self.joint_pos = self.cartpole.data.joint_pos
      self.joint_vel = self.cartpole.data.joint_vel

Scene Setup#

Scene setup is now done through the Cloner API and by specifying actor attributes in config objects. This eliminates the need to loop through the number of environments to set up the environments and avoids the need to set simulation parameters for actors in the task implementation.

IsaacGymEnvs

Isaac Lab

 
def create_sim(self):
    # set the up axis to be z-up given that assets are y-up by default
    self.up_axis = self.cfg["sim"]["up_axis"]

    self.sim = super().create_sim(self.device_id,
        self.graphics_device_id, self.physics_engine,
        self.sim_params)
    self._create_ground_plane()
    self._create_envs(self.num_envs,
        self.cfg["env"]['envSpacing'],
        int(np.sqrt(self.num_envs)))

def _create_ground_plane(self):
    plane_params = gymapi.PlaneParams()
    # set the normal force to be z dimension
    plane_params.normal = (gymapi.Vec3(0.0, 0.0, 1.0)
        if self.up_axis == 'z'
        else gymapi.Vec3(0.0, 1.0, 0.0))
    self.gym.add_ground(self.sim, plane_params)

def _create_envs(self, num_envs, spacing, num_per_row):
    # define plane on which environments are initialized
    lower = (gymapi.Vec3(0.5 * -spacing, -spacing, 0.0)
        if self.up_axis == 'z'
        else gymapi.Vec3(0.5 * -spacing, 0.0, -spacing))
    upper = gymapi.Vec3(0.5 * spacing, spacing, spacing)

    asset_root = os.path.join(os.path.dirname(
        os.path.abspath(__file__)), "../../assets")
    asset_file = "urdf/cartpole.urdf"

    if "asset" in self.cfg["env"]:
        asset_root = os.path.join(os.path.dirname(
            os.path.abspath(__file__)),
            self.cfg["env"]["asset"].get("assetRoot", asset_root))
        asset_file = self.cfg["env"]["asset"].get(
            "assetFileName", asset_file)

    asset_path = os.path.join(asset_root, asset_file)
    asset_root = os.path.dirname(asset_path)
    asset_file = os.path.basename(asset_path)

    asset_options = gymapi.AssetOptions()
    asset_options.fix_base_link = True
    cartpole_asset = self.gym.load_asset(self.sim,
        asset_root, asset_file, asset_options)
    self.num_dof = self.gym.get_asset_dof_count(
        cartpole_asset)

    pose = gymapi.Transform()
    if self.up_axis == 'z':
        pose.p.z = 2.0
        pose.r = gymapi.Quat(0.0, 0.0, 0.0, 1.0)
    else:
        pose.p.y = 2.0
        pose.r = gymapi.Quat(
            -np.sqrt(2)/2, 0.0, 0.0, np.sqrt(2)/2)

    self.cartpole_handles = []
    self.envs = []
    for i in range(self.num_envs):
        # create env instance
        env_ptr = self.gym.create_env(
            self.sim, lower, upper, num_per_row
        )
        cartpole_handle = self.gym.create_actor(
            env_ptr, cartpole_asset, pose,
            "cartpole", i, 1, 0)

        dof_props = self.gym.get_actor_dof_properties(
            env_ptr, cartpole_handle)
        dof_props['driveMode'][0] = gymapi.DOF_MODE_EFFORT
        dof_props['driveMode'][1] = gymapi.DOF_MODE_NONE
        dof_props['stiffness'][:] = 0.0
        dof_props['damping'][:] = 0.0
        self.gym.set_actor_dof_properties(env_ptr, c
            artpole_handle, dof_props)

        self.envs.append(env_ptr)
        self.cartpole_handles.append(cartpole_handle)

def _setup_scene(self):
    self.cartpole = Articulation(self.cfg.robot_cfg)
    # add ground plane
    spawn_ground_plane(prim_path="/World/ground",
        cfg=GroundPlaneCfg())
    # clone, filter, and replicate
    self.scene.clone_environments(
        copy_from_source=False)
    self.scene.filter_collisions(
        global_prim_paths=[])
    # add articulation to scene
    self.scene.articulations["cartpole"] = self.cartpole
    # add lights
    light_cfg = sim_utils.DomeLightCfg(
        intensity=2000.0, color=(0.75, 0.75, 0.75))
    light_cfg.func("/World/Light", light_cfg)

CARTPOLE_CFG = ArticulationCfg(
    spawn=sim_utils.UsdFileCfg(
        usd_path=f"{ISAACLAB_NUCLEUS_DIR}/.../cartpole.usd",
        rigid_props=sim_utils.RigidBodyPropertiesCfg(
            rigid_body_enabled=True,
            max_linear_velocity=1000.0,
            max_angular_velocity=1000.0,
            max_depenetration_velocity=100.0,
            enable_gyroscopic_forces=True,
        ),
        articulation_props=sim_utils.ArticulationRootPropertiesCfg(
            enabled_self_collisions=False,
            solver_position_iteration_count=4,
            solver_velocity_iteration_count=0,
            sleep_threshold=0.005,
            stabilization_threshold=0.001,
        ),
    ),
    init_state=ArticulationCfg.InitialStateCfg(
        pos=(0.0, 0.0, 2.0),
        joint_pos={"slider_to_cart": 0.0, "cart_to_pole": 0.0}
    ),
    actuators={
        "cart_actuator": ImplicitActuatorCfg(
            joint_names_expr=["slider_to_cart"],
            effort_limit=400.0,
            velocity_limit=100.0,
            stiffness=0.0,
            damping=10.0,
        ),
        "pole_actuator": ImplicitActuatorCfg(
            joint_names_expr=["cart_to_pole"], effort_limit=400.0,
            velocity_limit=100.0, stiffness=0.0, damping=0.0
        ),
    },
)

Pre and Post Physics Step#

In IsaacGymEnvs, due to limitations of the GPU APIs, observations had stale data when environments had to perform resets. This restriction has been eliminated in Isaac Lab, and thus, tasks follow the correct workflow of applying actions, stepping simulation, collecting states, computing dones, calculating rewards, performing resets, and finally computing observations. This workflow is done automatically by the framework such that a post_physics_step API is not required in the task. However, individual tasks can override the step() API to control the workflow.

IsaacGymEnvs

IsaacLab

 
def pre_physics_step(self, actions):
    actions_tensor = torch.zeros(
        self.num_envs * self.num_dof,
        device=self.device, dtype=torch.float)
    actions_tensor[::self.num_dof] = actions.to(
        self.device).squeeze() * self.max_push_effort
    forces = gymtorch.unwrap_tensor(actions_tensor)
    self.gym.set_dof_actuation_force_tensor(
        self.sim, forces)

def post_physics_step(self):
    self.progress_buf += 1

    env_ids = self.reset_buf.nonzero(
        as_tuple=False).squeeze(-1)
    if len(env_ids) > 0:
        self.reset_idx(env_ids)

    self.compute_observations()
    self.compute_reward()

def _pre_physics_step(self, actions: torch.Tensor) -> None:
    self.actions = self.action_scale * actions

def _apply_action(self) -> None:
    self.cartpole.set_joint_effort_target(
         self.actions, joint_ids=self._cart_dof_idx)

Dones and Resets#

In Isaac Lab, dones are computed in the _get_dones() method and should return two variables: resets and time_out. Tracking of the progress_buf has been moved to the base class and is now automatically incremented and reset by the framework. The progress_buf variable has also been renamed to episode_length_buf.

IsaacGymEnvs

Isaac Lab

 
def reset_idx(self, env_ids):
    positions = 0.2 * (torch.rand((len(env_ids), self.num_dof),
        device=self.device) - 0.5)
    velocities = 0.5 * (torch.rand((len(env_ids), self.num_dof),
        device=self.device) - 0.5)

    self.dof_pos[env_ids, :] = positions[:]
    self.dof_vel[env_ids, :] = velocities[:]

    env_ids_int32 = env_ids.to(dtype=torch.int32)
    self.gym.set_dof_state_tensor_indexed(self.sim,
        gymtorch.unwrap_tensor(self.dof_state),
        gymtorch.unwrap_tensor(env_ids_int32), len(env_ids_int32))
    self.reset_buf[env_ids] = 0
    self.progress_buf[env_ids] = 0

def _get_dones(self) -> tuple[torch.Tensor, torch.Tensor]:
    self.joint_pos = self.cartpole.data.joint_pos
    self.joint_vel = self.cartpole.data.joint_vel

    time_out = self.episode_length_buf >= self.max_episode_length - 1
    out_of_bounds = torch.any(torch.abs(
        self.joint_pos[:, self._cart_dof_idx]) > self.cfg.max_cart_pos,
        dim=1)
    out_of_bounds = out_of_bounds | torch.any(
        torch.abs(self.joint_pos[:, self._pole_dof_idx]) > math.pi / 2,
        dim=1)
    return out_of_bounds, time_out

def _reset_idx(self, env_ids: Sequence[int] | None):
    if env_ids is None:
        env_ids = self.cartpole._ALL_INDICES
    super()._reset_idx(env_ids)

    joint_pos = self.cartpole.data.default_joint_pos[env_ids]
    joint_pos[:, self._pole_dof_idx] += sample_uniform(
        self.cfg.initial_pole_angle_range[0] * math.pi,
        self.cfg.initial_pole_angle_range[1] * math.pi,
        joint_pos[:, self._pole_dof_idx].shape,
        joint_pos.device,
    )
    joint_vel = self.cartpole.data.default_joint_vel[env_ids]

    default_root_state = self.cartpole.data.default_root_state[env_ids]
    default_root_state[:, :3] += self.scene.env_origins[env_ids]

    self.joint_pos[env_ids] = joint_pos

    self.cartpole.write_root_link_pose_to_sim(
        default_root_state[:, :7], env_ids)
    self.cartpole.write_root_com_velocity_to_sim(
        default_root_state[:, 7:], env_ids)
    self.cartpole.write_joint_state_to_sim(
        joint_pos, joint_vel, None, env_ids)

Observations#

In Isaac Lab, the _get_observations() API should now return a dictionary containing the policy key with the observation buffer as the value. For asymmetric policies, the dictionary should also include a critic key that holds the state buffer.

IsaacGymEnvs

Isaac Lab

 
def compute_observations(self, env_ids=None):
    if env_ids is None:
        env_ids = np.arange(self.num_envs)

    self.gym.refresh_dof_state_tensor(self.sim)

    self.obs_buf[env_ids, 0] = self.dof_pos[env_ids, 0]
    self.obs_buf[env_ids, 1] = self.dof_vel[env_ids, 0]
    self.obs_buf[env_ids, 2] = self.dof_pos[env_ids, 1]
    self.obs_buf[env_ids, 3] = self.dof_vel[env_ids, 1]

    return self.obs_buf

def _get_observations(self) -> dict:
    obs = torch.cat(
        (
            self.joint_pos[:, self._pole_dof_idx[0]],
            self.joint_vel[:, self._pole_dof_idx[0]],
            self.joint_pos[:, self._cart_dof_idx[0]],
            self.joint_vel[:, self._cart_dof_idx[0]],
        ),
        dim=-1,
    )
    observations = {"policy": obs}
    return observations

Rewards#

In Isaac Lab, the reward method _get_rewards should return the reward buffer as a return value. Similar to IsaacGymEnvs, computations in the reward function can also be performed using pytorch jit by adding the @torch.jit.script annotation.

IsaacGymEnvs

Isaac Lab

 
def compute_reward(self):
    # retrieve environment observations from buffer
    pole_angle = self.obs_buf[:, 2]
    pole_vel = self.obs_buf[:, 3]
    cart_vel = self.obs_buf[:, 1]
    cart_pos = self.obs_buf[:, 0]

    self.rew_buf[:], self.reset_buf[:] = compute_cartpole_reward(
        pole_angle, pole_vel, cart_vel, cart_pos,
        self.reset_dist, self.reset_buf,
        self.progress_buf, self.max_episode_length
    )

@torch.jit.script
def compute_cartpole_reward(pole_angle, pole_vel,
                            cart_vel, cart_pos,
                            reset_dist, reset_buf,
                            progress_buf, max_episode_length):

    reward = (1.0 - pole_angle * pole_angle -
        0.01 * torch.abs(cart_vel) -
        0.005 * torch.abs(pole_vel))

    # adjust reward for reset agents
    reward = torch.where(torch.abs(cart_pos) > reset_dist,
        torch.ones_like(reward) * -2.0, reward)
    reward = torch.where(torch.abs(pole_angle) > np.pi / 2,
        torch.ones_like(reward) * -2.0, reward)

    reset = torch.where(torch.abs(cart_pos) > reset_dist,
        torch.ones_like(reset_buf), reset_buf)
    reset = torch.where(torch.abs(pole_angle) > np.pi / 2,
        torch.ones_like(reset_buf), reset_buf)
    reset = torch.where(progress_buf >= max_episode_length - 1,
        torch.ones_like(reset_buf), reset)

def _get_rewards(self) -> torch.Tensor:
    total_reward = compute_rewards(
        self.cfg.rew_scale_alive,
        self.cfg.rew_scale_terminated,
        self.cfg.rew_scale_pole_pos,
        self.cfg.rew_scale_cart_vel,
        self.cfg.rew_scale_pole_vel,
        self.joint_pos[:, self._pole_dof_idx[0]],
        self.joint_vel[:, self._pole_dof_idx[0]],
        self.joint_pos[:, self._cart_dof_idx[0]],
        self.joint_vel[:, self._cart_dof_idx[0]],
        self.reset_terminated,
    )
    return total_reward

@torch.jit.script
def compute_rewards(
    rew_scale_alive: float,
    rew_scale_terminated: float,
    rew_scale_pole_pos: float,
    rew_scale_cart_vel: float,
    rew_scale_pole_vel: float,
    pole_pos: torch.Tensor,
    pole_vel: torch.Tensor,
    cart_pos: torch.Tensor,
    cart_vel: torch.Tensor,
    reset_terminated: torch.Tensor,
):
    rew_alive = rew_scale_alive * (1.0 - reset_terminated.float())
    rew_termination = rew_scale_terminated * reset_terminated.float()
    rew_pole_pos = rew_scale_pole_pos * torch.sum(
        torch.square(pole_pos), dim=-1)
    rew_cart_vel = rew_scale_cart_vel * torch.sum(
        torch.abs(cart_vel), dim=-1)
    rew_pole_vel = rew_scale_pole_vel * torch.sum(
        torch.abs(pole_vel), dim=-1)
    total_reward = (rew_alive + rew_termination
                     + rew_pole_pos + rew_cart_vel + rew_pole_vel)
    return total_reward

Launching Training#

To launch a training in Isaac Lab, use the command:

python source/standalone/workflows/rl_games/train.py --task=Isaac-Cartpole-Direct-v0 --headless

Launching Inferencing#

To launch inferencing in Isaac Lab, use the command:

python source/standalone/workflows/rl_games/play.py --task=Isaac-Cartpole-Direct-v0 --num_envs=25 --checkpoint=<path/to/checkpoint>