Creating Custom Training Environment in IsaacLab via Extensions

As mentioned in the "Dive Into IsaacLab" article, the codebase of IsaacLab is significantly larger than other RL frameworks.

Case Study: A Dive Into IsaacLab

It provides a nice abstraction of configuration fragments using the Config class and the notion of Managers. However, this makes editing the environment and implementing custom training logic harder. Jumping across directory structures to access config fragments is common and annoying.

Fortunately, IsaacLab also supports Extensions, which allows us to isolate out the custom scripts into a separate repository. In this article, we will go over how to create such extension and achieve custom training logic.

Cloning the extension template

git clone https://github.com/isaac-sim/IsaacLabExtensionTemplate.git
cd ./IsaacLabExtensionTemplate/

Rename the template

We can rename the template to a custom extension name. Here, we will use g1_demo as an example:

python scripts/rename_template.py g1_demo

After executing this command, all the ext_template text inside this repo should be replace to g1_demo.

Configure the tasks

We start off by copying the scripts directly from upstream IsaacLab

There's a few difference between the scripts used in IsaacLab versus in the Extension.

velocity_env_cfg.py

# IsaacLab
import omni.isaac.lab_tasks.manager_based.locomotion.velocity.mdp as mdp

# Extension
import g1_demo.tasks.locomotion.velocity.mdp as mdp

train.py & play.py

# IsaacLab
import omni.isaac.lab_tasks  # noqa: F401

# Extension
import g1_demo.tasks  # noqa: F401