Skip to navigation
Isaac LabHow-to Guides

Creating a Direct Workflow RL Environment

View as Markdown

This tutorial is the source of truth for the isaaclab-randomizing-with-events agent skill (skills/user/domain-randomization-events/). When you change this page, update the skill so agent guidance stays in sync. See /source/developer-tools/agent_skills.

In addition to the envs.ManagerBasedRLEnv class, which encourages the use of configuration classes for more modular environments, the ~isaaclab.envs.DirectRLEnv class allows for more direct control in the scripting of environment.

Instead of using Manager classes for defining rewards and observations, the direct workflow tasks implement the full reward and observation functions directly in the task script. This allows for more control in the implementation of the methods, such as using pytorch jit features, and provides a less abstracted framework that makes it easier to find the various pieces of code.

In this tutorial, we will configure the cartpole environment using the direct workflow implementation to create a task for balancing the pole upright. We will declare its scene and implement actions, resets, rewards, and observations.

The Code

For this tutorial, we use the cartpole environment defined in isaaclab_tasks.core.cartpole module.

# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause
"""Direct-workflow cartpole balancing environment."""
from __future__ import annotations
from collections.abc import Sequence
from typing import TYPE_CHECKING
import torch
from isaaclab.envs import DirectRLEnv
from isaaclab.utils.math import sample_uniform, wrap_to_pi
if TYPE_CHECKING:
    from .cartpole_direct_env_cfg import CartpoleEnvCfg
class CartpoleEnv(DirectRLEnv):
    """Cartpole balancing environment driven by proprioceptive (joint-state) observations."""
    cfg: CartpoleEnvCfg
    def __init__(self, cfg: CartpoleEnvCfg, render_mode: str | None = None, **kwargs):
        super().__init__(cfg, render_mode, **kwargs)
        self.cartpole = self.scene["cartpole"]
        cart_dof_idx, _ = self.cartpole.find_joints(self.cfg.cart_dof_name)
        pole_dof_idx, _ = self.cartpole.find_joints(self.cfg.pole_dof_name)
        # device indices avoid per-step host uploads
        self._cart_dof_idx = torch.tensor(cart_dof_idx, device=self.device)
        self._pole_dof_idx = torch.tensor(pole_dof_idx, device=self.device)
        self.action_scale = self.cfg.action_scale
        self.joint_pos = self.cartpole.data.joint_pos.torch
        self.joint_vel = self.cartpole.data.joint_vel.torch
    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_index(target=self.actions, joint_ids=self._cart_dof_idx)
    def _get_observations(self) -> dict:
        joint_pos_rel = self.joint_pos - self.cartpole.data.default_joint_pos.torch
        joint_vel_rel = self.joint_vel - self.cartpole.data.default_joint_vel.torch
        obs = torch.cat(
            (
                joint_pos_rel[:, self._cart_dof_idx],
                joint_pos_rel[:, self._pole_dof_idx],
                joint_vel_rel[:, self._cart_dof_idx],
                joint_vel_rel[:, self._pole_dof_idx],
            ),
            dim=-1,
        )
        return {"policy": obs}
    def _get_rewards(self) -> torch.Tensor:
        return 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],
            self.joint_vel[:, self._pole_dof_idx],
            self.joint_vel[:, self._cart_dof_idx],
            self.reset_terminated,
            self.step_dt,
        )
    def _get_dones(self) -> tuple[torch.Tensor, torch.Tensor]:
        self.joint_pos = self.cartpole.data.joint_pos.torch
        self.joint_vel = self.cartpole.data.joint_vel.torch
        time_out = self.episode_length_buf >= self.max_episode_length
        out_of_bounds = torch.any(torch.abs(self.joint_pos[:, self._cart_dof_idx]) > self.cfg.max_cart_pos, 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
        # log the survival success rate before resetting (survived = timed out without terminating early)
        survived = self.reset_time_outs[env_ids].float()
        # no .item(): avoids a sync on every reset
        self.extras.setdefault("log", {})["Metrics/success_rate"] = survived.mean()
        super()._reset_idx(env_ids)
        joint_pos = self.cartpole.data.default_joint_pos.torch[env_ids].clone()
        joint_pos[:, self._cart_dof_idx] += sample_uniform(
            self.cfg.initial_cart_position_range[0],
            self.cfg.initial_cart_position_range[1],
            joint_pos[:, self._cart_dof_idx].shape,
            joint_pos.device,
        )
        joint_vel = self.cartpole.data.default_joint_vel.torch[env_ids].clone()
        joint_vel[:, self._cart_dof_idx] += sample_uniform(
            self.cfg.initial_cart_velocity_range[0],
            self.cfg.initial_cart_velocity_range[1],
            joint_vel[:, self._cart_dof_idx].shape,
            joint_vel.device,
        )
        joint_pos[:, self._pole_dof_idx] += sample_uniform(
            self.cfg.initial_pole_angle_range[0],
            self.cfg.initial_pole_angle_range[1],
            joint_pos[:, self._pole_dof_idx].shape,
            joint_pos.device,
        )
        joint_vel[:, self._pole_dof_idx] += sample_uniform(
            self.cfg.initial_pole_velocity_range[0],
            self.cfg.initial_pole_velocity_range[1],
            joint_vel[:, self._pole_dof_idx].shape,
            joint_vel.device,
        )
        # clamp the sampled state to the joint limits (matches the manager-based reset_joints_by_offset)
        joint_pos_limits = self.cartpole.data.soft_joint_pos_limits.torch[env_ids]
        joint_pos = joint_pos.clamp_(joint_pos_limits[..., 0], joint_pos_limits[..., 1])
        joint_vel_limits = self.cartpole.data.soft_joint_vel_limits.torch[env_ids]
        joint_vel = joint_vel.clamp_(-joint_vel_limits, joint_vel_limits)
        default_root_pose = self.cartpole.data.default_root_pose.torch[env_ids].clone()
        default_root_pose[:, :3] += self.scene.env_origins[env_ids]
        default_root_vel = self.cartpole.data.default_root_vel.torch[env_ids].clone()
        self.joint_pos[env_ids] = joint_pos
        self.joint_vel[env_ids] = joint_vel
        self.cartpole.write_root_pose_to_sim_index(root_pose=default_root_pose, env_ids=env_ids)
        self.cartpole.write_root_velocity_to_sim_index(root_velocity=default_root_vel, env_ids=env_ids)
        self.cartpole.write_joint_position_to_sim_index(position=joint_pos, env_ids=env_ids)
        self.cartpole.write_joint_velocity_to_sim_index(velocity=joint_vel, env_ids=env_ids)
@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_vel: torch.Tensor,
    reset_terminated: torch.Tensor,
    step_dt: float,
) -> torch.Tensor:
    pole_pos = wrap_to_pi(pole_pos)
    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) * step_dt
    return total_reward

The Code Explained

Similar to the manager-based environments, a configuration class is defined for the task to hold settings for the simulation parameters, the scene, the actors, and the task. With the direct workflow implementation, the envs.DirectRLEnvCfg class is used as the base class for configurations. Since the direct workflow implementation does not use Action and Observation managers, the task config should define the number of actions and observations for the environment.

@configclass
class CartpoleEnvCfg(DirectRLEnvCfg):
   ...
   action_space = 1
   observation_space = 4
   state_space = 0

The config class can also be used to define task-specific attributes, such as scaling for reward terms and thresholds for reset conditions.

@configclass
class CartpoleEnvCfg(DirectRLEnvCfg):
   ...
   # 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

When creating a new environment, the code should define a new class that inherits from ~isaaclab.envs.DirectRLEnv.

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

The class can also hold class variables that are accessible by all functions in the class, including functions for applying actions, computing resets, rewards, and observations.

Scene Creation

Direct and manager-based environments use the same declarative scene lifecycle. Define the assets and sensors on an ~isaaclab.scene.InteractiveSceneCfg subclass:

@configclass
class CartpoleSceneCfg(InteractiveSceneCfg):
    """Cartpole assets constructed and cloned as one scene."""
    ground = AssetBaseCfg(prim_path="/World/ground", spawn=sim_utils.GroundPlaneCfg())
    cartpole: ArticulationCfg = replace(CARTPOLE_CFG, prim_path="{ENV_REGEX_NS}/Robot")
    light = AssetBaseCfg(
        prim_path="/World/Light",
        spawn=sim_utils.DistantLightCfg(intensity=2000.0),
        init_state=AssetBaseCfg.InitialStateCfg(rot=LIGHT_ORIENTATION),
    )

Assign that configuration to the task’s scene field. ~isaaclab.envs.DirectRLEnv constructs the scene and completes its one clone lifecycle before the task constructor resumes. Task code can then keep a convenient reference to a declared entity:

self.cartpole = self.scene["cartpole"]

Defining Rewards

Reward function should be defined in the _get_rewards(self) API, which returns the reward buffer as a return value. Within this function, the task is free to implement the logic of the reward function. In this example, we implement a Pytorch JIT function that computes the various components of the reward function.

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

Defining Observations

The observation buffer should be computed in the _get_observations(self) function, which constructs the observation buffer for the environment. At the end of this API, a dictionary should be returned that contains policy as the key, and the full observation buffer as the value. For asymmetric policies, the dictionary should also include the key critic and the states buffer as the value.

def _get_observations(self) -> dict:
    joint_pos_rel = self.joint_pos - self.cartpole.data.default_joint_pos.torch
    joint_vel_rel = self.joint_vel - self.cartpole.data.default_joint_vel.torch
    obs = torch.cat(
        (
            joint_pos_rel[:, self._cart_dof_idx],
            joint_pos_rel[:, self._pole_dof_idx],
            joint_vel_rel[:, self._cart_dof_idx],
            joint_vel_rel[:, self._pole_dof_idx],
        ),
        dim=-1,
    )
    return {"policy": obs}

Computing Dones and Performing Resets

Populating the dones buffer should be done in the _get_dones(self) method. This method is free to implement logic that computes which environments would need to be reset and which environments have reached the episode length limit. Both results should be returned by the _get_dones(self) function, in the form of a tuple of boolean tensors.

def _get_dones(self) -> tuple[torch.Tensor, torch.Tensor]:
    self.joint_pos = self.cartpole.data.joint_pos.torch
    self.joint_vel = self.cartpole.data.joint_vel.torch
    time_out = self.episode_length_buf >= self.max_episode_length
    out_of_bounds = torch.any(torch.abs(self.joint_pos[:, self._cart_dof_idx]) > self.cfg.max_cart_pos, dim=1)
    return out_of_bounds, time_out

Once the indices for environments requiring reset have been computed, the _reset_idx(self, env_ids) function performs the reset operations on those environments. Within this function, new states for the environments requiring reset should be set directly into simulation.

def _reset_idx(self, env_ids: Sequence[int] | None):
    if env_ids is None:
        env_ids = self.cartpole._ALL_INDICES
    # log the survival success rate before resetting (survived = timed out without terminating early)
    survived = self.reset_time_outs[env_ids].float()
    # no .item(): avoids a sync on every reset
    self.extras.setdefault("log", {})["Metrics/success_rate"] = survived.mean()
    super()._reset_idx(env_ids)
    joint_pos = self.cartpole.data.default_joint_pos.torch[env_ids].clone()
    joint_pos[:, self._cart_dof_idx] += sample_uniform(
        self.cfg.initial_cart_position_range[0],
        self.cfg.initial_cart_position_range[1],
        joint_pos[:, self._cart_dof_idx].shape,
        joint_pos.device,
    )
    joint_vel = self.cartpole.data.default_joint_vel.torch[env_ids].clone()
    joint_vel[:, self._cart_dof_idx] += sample_uniform(
        self.cfg.initial_cart_velocity_range[0],
        self.cfg.initial_cart_velocity_range[1],
        joint_vel[:, self._cart_dof_idx].shape,
        joint_vel.device,
    )
    joint_pos[:, self._pole_dof_idx] += sample_uniform(
        self.cfg.initial_pole_angle_range[0],
        self.cfg.initial_pole_angle_range[1],
        joint_pos[:, self._pole_dof_idx].shape,
        joint_pos.device,
    )
    joint_vel[:, self._pole_dof_idx] += sample_uniform(
        self.cfg.initial_pole_velocity_range[0],
        self.cfg.initial_pole_velocity_range[1],
        joint_vel[:, self._pole_dof_idx].shape,
        joint_vel.device,
    )
    # clamp the sampled state to the joint limits (matches the manager-based reset_joints_by_offset)
    joint_pos_limits = self.cartpole.data.soft_joint_pos_limits.torch[env_ids]
    joint_pos = joint_pos.clamp_(joint_pos_limits[..., 0], joint_pos_limits[..., 1])
    joint_vel_limits = self.cartpole.data.soft_joint_vel_limits.torch[env_ids]
    joint_vel = joint_vel.clamp_(-joint_vel_limits, joint_vel_limits)
    default_root_pose = self.cartpole.data.default_root_pose.torch[env_ids].clone()
    default_root_pose[:, :3] += self.scene.env_origins[env_ids]
    default_root_vel = self.cartpole.data.default_root_vel.torch[env_ids].clone()
    self.joint_pos[env_ids] = joint_pos
    self.joint_vel[env_ids] = joint_vel
    self.cartpole.write_root_pose_to_sim_index(root_pose=default_root_pose, env_ids=env_ids)
    self.cartpole.write_root_velocity_to_sim_index(root_velocity=default_root_vel, env_ids=env_ids)
    self.cartpole.write_joint_position_to_sim_index(position=joint_pos, env_ids=env_ids)
    self.cartpole.write_joint_velocity_to_sim_index(velocity=joint_vel, env_ids=env_ids)

Applying Actions

There are two APIs that are designed for working with actions. The _pre_physics_step(self, actions) takes in actions from the policy as an argument and is called once per RL step, prior to taking any physics steps. This function can be used to process the actions buffer from the policy and cache the data in a class variable for the environment.

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

The _apply_action(self) API is called decimation number of times for each RL step, prior to taking each physics step. This provides more flexibility for environments where actions should be applied for each physics step.

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

The Code Execution

To run training for the direct workflow Cartpole environment, we can use the following command:

uv run --extra rl-games isaaclab train --rl_library rl_games --task=Isaac-Cartpole-Direct
result of train.py

All direct workflow tasks have the suffix -Direct added to the task name to differentiate the implementation style.

Domain Randomization

In the direct workflow, domain randomization configuration uses the ~isaaclab.utils.configclass module to specify a configuration class consisting of ~managers.EventTermCfg variables.

Below is an example of a configuration class for domain randomization:

@configclass
class EventCfg:
  robot_physics_material = EventTerm(
      func=mdp.randomize_rigid_body_material,
      mode="reset",
      params={
          "asset_cfg": SceneEntityCfg("robot", body_names=".*"),
          "static_friction_range": (0.7, 1.3),
          "dynamic_friction_range": (1.0, 1.0),
          "restitution_range": (1.0, 1.0),
          "num_buckets": 250,
      },
  )
  robot_joint_stiffness_and_damping = EventTerm(
      func=mdp.randomize_actuator_gains,
      mode="reset",
      params={
          "asset_cfg": SceneEntityCfg("robot", joint_names=".*"),
          "stiffness_distribution_params": (0.75, 1.5),
          "damping_distribution_params": (0.3, 3.0),
          "operation": "scale",
          "distribution": "log_uniform",
      },
  )
  reset_gravity = EventTerm(
      func=mdp.randomize_physics_scene_gravity,
      mode="interval",
      is_global_time=True,
      interval_range_s=(36.0, 36.0),  # time_s = num_steps * (decimation * dt)
      params={
          "gravity_distribution_params": ([0.0, 0.0, 0.0], [0.0, 0.0, 0.4]),
          "operation": "add",
          "distribution": "gaussian",
      },
  )

Each EventTerm object is of the ~managers.EventTermCfg class and takes in a func parameter for specifying the function to call during randomization, a mode parameter, which can be startup, reset or interval. The params dictionary should provide the necessary arguments to the function that is specified in the func parameter. Functions specified as func for the EventTerm can be found in the ~envs.mdp.events module.

Note that as part of the "asset_cfg": SceneEntityCfg("robot", body_names=".*") parameter, the name of the actor "robot" is provided, along with the body or joint names specified as a regex expression, which will be the actors and bodies/joints that will have randomization applied.

Once the configclass for the randomization terms have been set up, the class must be added to the base config class for the task and be assigned to the variable events.

@configclass
class MyTaskConfig:
  events: EventCfg = EventCfg()

Action and Observation Noise

Actions and observation noise can also be added using the ~utils.configclass module. Action and observation noise configs must be added to the main task config using the action_noise_model and observation_noise_model variables:

@configclass
class MyTaskConfig:
    # at every time-step add gaussian noise + bias. The bias is a gaussian sampled at reset
    action_noise_model: NoiseModelWithAdditiveBiasCfg = NoiseModelWithAdditiveBiasCfg(
      noise_cfg=GaussianNoiseCfg(mean=0.0, std=0.05, operation="add"),
      bias_noise_cfg=GaussianNoiseCfg(mean=0.0, std=0.015, operation="abs"),
    )
    # at every time-step add gaussian noise + bias. The bias is a gaussian sampled at reset
    observation_noise_model: NoiseModelWithAdditiveBiasCfg = NoiseModelWithAdditiveBiasCfg(
      noise_cfg=GaussianNoiseCfg(mean=0.0, std=0.002, operation="add"),
      bias_noise_cfg=GaussianNoiseCfg(mean=0.0, std=0.0001, operation="abs"),
    )

~.utils.noise.NoiseModelWithAdditiveBiasCfg can be used to sample both uncorrelated noise per step as well as correlated noise that is re-sampled at reset time.

The noise_cfg term specifies the Gaussian distribution that will be sampled at each step for all environments. This noise will be added to the corresponding actions and observations buffers at every step.

The bias_noise_cfg term specifies the Gaussian distribution for the correlated noise that will be sampled at reset time for the environments being reset. The same noise will be applied each step for the remaining of the episode for the environments and resampled at the next reset.

If only per-step noise is desired, ~utils.noise.GaussianNoiseCfg can be used to specify an additive Gaussian distribution that adds the sampled noise to the input buffer.

@configclass
class MyTaskConfig:
  action_noise_model: GaussianNoiseCfg = GaussianNoiseCfg(mean=0.0, std=0.05, operation="add")

In this tutorial, we learnt how to create a direct workflow task environment for reinforcement learning. We declared the scene in configuration and extended the base environment with actions, dones, reset, reward, and observation functions.

While it is possible to manually create an instance of ~isaaclab.envs.DirectRLEnv class for a desired task, this is not scalable as it requires specialized scripts for each task. Thus, we exploit the gymnasium.make function to create the environment with the gym interface. We will learn how to do this in the next tutorial.