isaaclab_contrib.rl

isaaclab_contrib.rl#

Reinforcement learning extensions contributed by the community.

Submodules#

isaaclab_contrib.rl.rlinf#

RLinf integration for IsaacLab.

This package provides the extension mechanism for integrating IsaacLab tasks with RLinf’s distributed RL training framework for VLA models like GR00T.

Extension Module#

The extension module (extension.py) is loaded by RLinf via the RLINF_EXT_MODULE environment variable and handles:

  1. Registering IsaacLab tasks into RLinf’s REGISTER_ISAACLAB_ENVS

  2. Registering GR00T obs/action converters

  3. Patching GR00T get_model for custom embodiments

Usage:
export RLINF_EXT_MODULE="isaaclab_contrib.rl.rlinf.extension"
export RLINF_CONFIG_FILE="/path/to/config.yaml"

Note

The extension module requires the external rlinf package and cannot be introspected at documentation build time. The API is described textually below.

Extension Module#

The extension module (isaaclab_contrib.rl.rlinf.extension) is loaded by RLinf’s worker framework via the RLINF_EXT_MODULE environment variable. It is not imported directly by user code.

Setup:

export RLINF_EXT_MODULE="isaaclab_contrib.rl.rlinf.extension"
export RLINF_CONFIG_FILE="/path/to/config.yaml"

Public entry point:

  • register() – Called by RLinf’s worker to perform all setup. It:

    1. Registers GR00T observation and action converters.

    2. Patches GR00T’s get_model for custom embodiment tags.

    3. Registers IsaacLab tasks into RLinf’s REGISTER_ISAACLAB_ENVS registry.

Expected YAML configuration (under env.train.isaaclab):

env:
  train:
    isaaclab: &isaaclab_config
      task_description: "Assemble trocar with dual-arm robot"
      main_images: "front_camera"
      extra_view_images: ["left_wrist_camera", "right_wrist_camera"]
      states:
        - key: "robot_joint_state"
          slice: [15, 29]
      gr00t_mapping:
        video:
          main_images: "video.room_view"
      action_mapping:
        prefix_pad: 15
  eval:
    isaaclab: *isaaclab_config  # Reuse via YAML anchor

Task IDs are read automatically from env.train.init_params.id and env.eval.init_params.id in the YAML config.