RobRAN Code:

A Unified Robotics Framework for Reinforcement Learning-Based Autonomous Navigation.

This tutorial shows hot to create a robotic system, linking it to tasks, and training a Reinforcement Learning (RL) agent using the RANS version of Isaac Lab framework. As an example, we use the robot named floating_platform.

Website

• https://snt-spacer.github.io/RoboRAN-Website/

Deployment To Real Robots Code

• https://github.com/snt-spacer/RoboRAN-deploy-to-robot

BibTex

@article{
anonymous2025roboran,
title={Robo{RAN}: A Unified Robotics Framework for Reinforcement Learning-Based Autonomous Navigation},
author={Anonymous},
journal={Submitted to Transactions on Machine Learning Research},
year={2025},
url={https://openreview.net/forum?id=0wDbhLeMj9},
}

Overview of the Example Files and Structure

Main Folders and Their Roles

• RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/rans/robots/: Contains the robot configuration and class descriptions. For example, floating_platform.py specifies the floating_platform robot's properties.

• RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/direct/ROBOT_NAME: Contains task descriptions for robot-task pairs. Tasks are registered as Gym environments with unique IDs in the __init__.py file.

• Subdirectories of rans:

  • robots: Contains specific robot class definitions.
  • robots_cfg: Contains robot-specific configurations.
  • tasks: Implements various tasks for the robots, like navigation or velocity tracking.
  • tasks_cfg: Contains task-specific configuration files.
  • utils: Provides utility functions like generation of unique per-environment RNG, functionalities to evaluate and plot the results of testing a trained model for a specific robot-task pair.

Step 1: Configuring the Robot

Example File: floating_platform.py

Location: RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/rans/robots/floating_platform.py

This file defines the robot configuration using the FLOATING_PLATFORM_CFG object. Key aspects include:

• USD Path: Specifies the 3D model file.
• Physical Properties: Enables simulation features like gravity and velocity constraints.
• Initial State: Sets default positions and joint values.

Example:

FLOATING_PLATFORM_CFG = ArticulationCfg(
    spawn=sim_utils.UsdFileCfg(
        usd_path=f"{REPO_ROOT_PATH}/floating_platform.usd",
        rigid_props=sim_utils.RigidBodyPropertiesCfg(disable_gravity=True),
    ),
    init_state=ArticulationCfg.InitialStateCfg(pos=(0.0, 0.0, 0.5)),
)

Step 2: Defining Tasks

Example Folder: tasks/

Location: RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/rans/tasks/

Tasks are implemented here. For example:

• File: go_to_position.py describes the "Go To Position" task.
• Core Task Logic: Uses the TaskCore base class to define observation space, rewards, and task dynamics.

Example:

class GoToPositionTask(TaskCore):
    def get_observations(self):
        # Observation logic
        pass

    def compute_rewards(self):
        # Reward logic
        pass

Step 3: Configuring Task Environments

Example File: floating_platform_go_to_position.py

Location: RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/direct/floating_platform/floating_platform_go_to_position.py

This file links the robot and task configurations to create an RL environment.

Example:

@configclass
class FloatingPlatformGoToPositionEnvCfg(DirectRLEnvCfg):
    scene = InteractiveSceneCfg(num_envs=4096)
    robot_cfg = FloatingPlatformRobotCfg()
    task_cfg = GoToPositionCfg()
    episode_length_s = 40.0

Step 4: Registering Gym Environments

Example File: __init__.py

Location: RobRAN-Code/source/isaaclab_tasks/isaaclab_tasks/direct/floating_platform/__init__.py

Gym environments are registered with unique IDs, linking them to specific task and robot configurations.

Example registration:

gym.register(
    id="Isaac-FloatingPlatform-GoToPosition-Direct-v0",
    entry_point=f"{__name__}.floating_platform_go_to_position_env:FloatingPlatformGoToPositionEnv",
    disable_env_checker=True,
    kwargs={
        "env_cfg_entry_point": f"{__name__}.floating_platform_go_to_position_env:FloatingPlatformGoToPositionEnvCfg",
        "rl_games_cfg_entry_point": f"{agents.__name__}:rl_games_ppo_cfg.yaml",
        "rsl_rl_cfg_entry_point": f"{agents.__name__}.rsl_rl_ppo_cfg:FloatingPlatformPPORunnerCfg",
        "skrl_cfg_entry_point": f"{agents.__name__}:skrl_ppo_cfg.yaml",
        "sb3_cfg_entry_point": f"{agents.__name__}:sb3_ppo_cfg.yaml",
    },
)

Step 5: Training an RL Agent

RL Libraries

Compatible libraries include rl_games, skrl, and Stable-Baselines3. Example configuration files are referenced in the __init__.py.

Training Command

Start the docker:

./docker/container.py start
./docker/container.py enter

Use the environment ID (e.g., "Isaac-FloatingPlatform-GoToPosition-Direct-v0") to start training.

Example command to train a policy with rl_games:

 ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py --task Isaac-FloatingPlatform-GoToPosition-Direct-v0 --num_envs 4096 --headless

Example command to play a trained policy with rl_games:

 ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/play.py --task Isaac-FloatingPlatform-GoToPosition-Direct-v0 --num_envs 16

Alternatively, to easily switch between robot and tasks, one can use the Single environment:

./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py --task Isaac-RANS-Single-v0 --num_envs 4096 --headless env.robot_name=FloatingPlatform env.task_name=GoToPosition

In the previous command, the env.robot_name is used to select which robot should be used, and the env.task_name is used to select the task that should be loaded. These names relate to the ones given inside the factories in these files:

source/isaaclab_tasks/isaaclab_tasks/rans/robots_cfg/__init__.py
source/isaaclab_tasks/isaaclab_tasks/rans/robots/__init__.py
source/isaaclab_tasks/isaaclab_tasks/rans/tasks_cfg/__init__.py
source/isaaclab_tasks/isaaclab_tasks/rans/tasks/__init__.py