Modifying an existing Direct RL Environment

Having learnt how to create a task in Creating a Direct Workflow RL Environment, register it in Registering an Environment, and train it in Training with an RL Agent, we will now look at how to make minor modifications to an existing task.

Sometimes it is necessary to create, due to complexity or variations from existing examples, tasks from scratch. However, in certain situations, it is possible to start from the existing code and introduce minor changes, one by one, to transform them according to our needs.

In this tutorial, we will make minor modifications to the direct workflow Humanoid task to change the simple humanoid model to the Unitree H1 humanoid robot without affecting the original code.

The Base Code

For this tutorial, we start from the direct workflow Humanoid environment defined in isaaclab_tasks.core.locomotion.humanoid 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 Humanoid locomotion environment."""
from __future__ import annotations
from ..locomotion_direct_env import LocomotionDirectEnv
from .humanoid_direct_env_cfg import HumanoidEnvCfg
class HumanoidEnv(LocomotionDirectEnv):
    """Direct-workflow Humanoid locomotion environment.
    The behavior is fully defined by `LocomotionDirectEnv` and `HumanoidEnvCfg`.
    """
    cfg: HumanoidEnvCfg

The Changes Explained

Deriving and registering a new task

To avoid modifying the existing task, create an h1_env.py file in the Isaac Lab project source/isaaclab_tasks/isaaclab_tasks/core/locomotion/humanoid folder. The new classes below inherit the existing environment and configuration, changing only the robot-specific data.

Register the derived task under the name Isaac-H1-Direct-v0. To do this, we modify the __init__.py file in the same working folder and add the following entry. Refer to the Registering an Environment tutorial for more details about environment registrations.

from .h1_env import H1Env, H1EnvCfg

gym.register(
    id="Isaac-H1-Direct-v0",
    entry_point="isaaclab_tasks.core.locomotion.humanoid:H1Env",
    disable_env_checker=True,
    kwargs={
        "env_cfg_entry_point": H1EnvCfg,
        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_direct_ppo_cfg.yaml",
        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:HumanoidDirectPPORunnerCfg",
        "default_agent": "rsl_rl",
        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_direct_ppo_cfg.yaml",
    },
)

Changing the robot

The H1SceneCfg class (in the newly created h1_env.py file) declares the assets that the environment constructs. In this example, its robot property holds the target articulation configuration.

Since the Unitree H1 robot is included in the Isaac Lab assets extension (isaaclab_assets) we can just import it and do the replacement directly (under the H1SceneCfg.robot property), as shown below. Note that we also need to modify the joint_gears property as it holds robot-specific configuration values.

from isaaclab.assets import ArticulationCfg
from isaaclab.utils import configclass, replace
from isaaclab_assets import H1_CFG
from isaaclab_tasks.core.locomotion.humanoid.humanoid_direct_env import HumanoidEnv
from isaaclab_tasks.core.locomotion.humanoid.humanoid_direct_env_cfg import HumanoidDirectSceneCfg, HumanoidEnvCfg

@configclass
class H1SceneCfg(HumanoidDirectSceneCfg):
    robot: ArticulationCfg = replace(H1_CFG, prim_path="{ENV_REGEX_NS}/Robot")

The robot changed, and with it the number of joints to control or the number of rigid bodies that compose the articulation, for example. Therefore, define the derived environment configuration and adjust the values that depend on the robot, such as the action and observation spaces. The environment implementation itself remains unchanged.

The H1 has 19 controlled joints. The inherited observation function combines 12 root and target values, 19 joint positions, 19 joint velocities, 19 actions, and 12 wrench values from its two feet, for 81 values. Set feet_body_names to match H1’s ankle links so the joint-wrench sensor selects the correct bodies instead of the base Humanoid’s foot names.

@configclass
class H1EnvCfg(HumanoidEnvCfg):
    action_space = 19
    observation_space = 81
    scene: H1SceneCfg = H1SceneCfg(num_envs=4096, env_spacing=5.0, replicate_physics=True, clone_in_fabric=True)
    # the gears are keyed by joint name expression, so they are independent of the joint ordering
    joint_gears: dict[str, float] = {".*": 50.0}
    feet_body_names: list[str] = [".*ankle_link"]
class H1Env(HumanoidEnv):
    cfg: H1EnvCfg

The Code Execution

After the minor modification has been done, and similar to the previous tutorial, we can train on the task using one of the available RL workflows for such task.

uv run --extra rl-games isaaclab train --rl_library rl_games --task Isaac-H1-Direct-v0

When the training is finished, we can visualize the result with the following command. To stop the simulation, you can either close the window, or press Ctrl+C in the terminal where you started the simulation.

uv run --extra rl-games isaaclab play --rl_library rl_games --task Isaac-H1-Direct-v0 --num_envs 64 --viz kit

In this tutorial, we learnt how to make minor modifications to an existing environment without affecting the original code.

It is important to note, however, that while the changes to be made may be small, they may not always work on the first try, as there may be deeper dependencies on the original assets in the environment being modified. In these cases, it is advisable to analyze the code of the available examples in detail in order to make an appropriate adjustment.