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

This is a landing page for all the projects related to RobRAN. Below there's the links to the code of each project:

Main Code (TMLR)

• https://github.com/snt-spacer/RoboRAN-Code

Deployment To Real Robots Code

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

@article{
el-hariry2025roboran,
title={Robo{RAN}: A Unified Robotics Framework for Reinforcement Learning-Based Autonomous Navigation},
author={Matteo El-Hariry and Antoine Richard and Ricard Marsal and Luis Felipe Wolf Batista and Matthieu Geist and C{\'e}dric Pradalier and Miguel Olivares-Mendez},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=0wDbhLeMj9},
note={}
}

RobRAN is a multi-domain reinforcement learning benchmark designed for robotic navigation tasks in terrestrial, aquatic, and space environments. Built on IsaacLab, our framework enables:

✅ Fair comparisons across different robots and mobility systems ✅ Scalable training pipelines for reinforcement learning agents ✅ Sim-to-real transfer validation on physical robots

Turtlebot 2	Kingfisher	Floating platform

• Diverse Navigation Tasks: GoToPosition, GoToPose, GoThroughPositions, TrackVelocities, and more.
• Cross-Domain Evaluation: Supports thruster-based platforms, wheeled robots, and water-based propulsion.
• Unified Task Definitions: Standardized observation space, reward structures, and evaluation metrics.
• Efficient Simulation: GPU-accelerated rollouts via IsaacLab for rapid RL training.
• Real-World Validation: Policies successfully deployed on a Floating Platform, Kingfisher, and Turtlebot2.

Code lives in this anonymous repo:

git clone https://anonymous.4open.science/r/RobRAN-Code-E08E/README.md
cd RobRAN-Code
./docker/container.py start
./docker/container.py enter

./isaaclab.sh -p scripts/reinforcement_learning/<isaac_lab_rl_framework>/train.py --task=Isaac-RANS-Single-v0 --headless env.robot_name=<robot_name> env.task_name=<task>

• GoToPosition
• GoToPose
• GoThroughPositions
• TrackVelocities

Play trained models

./isaaclab.sh -p scripts/reinforcement_learning/<isaac_lab_rl_framework>/play.py --task=Isaac-RANS-Single-v0 --num_envs=32 env.robot_name=<robot_name> env.task_name=<task> --checkpoint=<path_to_pt_model>

Evaluation & Metrics

./isaaclab.sh -p scripts/reinforcement_learning/run_all_evals.py

The table below summarizes the performance of policies trained with skrl and rl_games on shared navigation tasks.

This table compares performance across tasks using PPO from two RL libraries: skrl and rl_games. While both show strong convergence, some variations emerge, particularly in heading control and velocity tracking. These differences likely stem from implementation details (e.g., optimizer behavior, action noise, or learning rate schedules). Despite these, both frameworks achieve high success rates and consistent trends, confirming that the benchmark stack is stable and the results are reproducible across PPO variants.

You can download all the trained models from this link.

Turtlebot 2	Kingfisher	Floating platform