RobRAN Website

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 (MTLR Submission)

• https://anonymous.4open.science/r/RobRAN-Code-E08E/

Deployment To Real Robots Code

• https://anonymous.4open.science/r/RobRAN-Deploynment-E08E/

Overview

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

🎥 Real-world deployments

Turtlebot 2	Kingfisher	Floating platform

Features

• 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.

🚧 Installation

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

Reproducibility

🧠 Training pipelines for all tasks and robots

./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>

Robots

Land	Water	Space
Jetbot	Kingfisher	FloatingPlatform
Leatherback
Turtlebot2

Tasks

• GoToPosition
• GoToPose
• GoThroughPositions
• TrackVelocities

Note

The paper was tested using SKRL and RL_Games for the isaac_lab_rl_framework.

PPO Hyperparameters

Parameter	Value
Rollouts	32
Learning Epochs	8
Mini Batches	8
Discount Factor	0.99
Lambda	0.95
Learning Rate	5.0e-04
KL Threshold	0.016
Epochs	1000
Network size	32x32

🧪 Evaluation and visualization

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

Performance Comparison Across RL Frameworks

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

Task	Robot	Success	Final Dist Err	Time to Target	Ctrl Var	Heading Err	Goals Reached
GoThroughPositions
	FloatingPlatform (skrl)	1.000	2.346	65.180	0.318	—	13.565
	FloatingPlatform (rl_games)	1.000	2.697	66.640	0.373	—	14.025
	Kingfisher (skrl)	1.000	2.414	93.290	0.430	—	10.702
	Kingfisher (rl_games)	1.000	3.525	67.050	0.092	—	14.716
	Turtlebot2 (skrl)	1.000	1.789	101.500	0.133	—	11.006
	Turtlebot2 (rl_games)	1.000	1.861	84.170	0.052	—	10.835
GoToPosition
	FloatingPlatform (skrl)	0.994	0.050	92.380	0.620	—	—
	FloatingPlatform (rl_games)	0.995	0.035	91.830	0.676	—	—
	Kingfisher (skrl)	0.589	1.063	176.110	0.750	—	—
	Kingfisher (rl_games)	0.998	0.023	90.040	0.112	—	—
	Turtlebot2 (skrl)	0.986	0.069	92.600	0.433	—	—
	Turtlebot2 (rl_games)	0.979	0.066	99.200	0.063	—	—
GoToPose
	FloatingPlatform (skrl)	0.993	0.024	92.370	0.688	0.783	—
	FloatingPlatform (rl_games)	0.979	0.035	88.710	0.754	0.801	—
	Turtlebot2 (skrl)	0.836	0.145	131.490	0.629	4.389	—
	Turtlebot2 (rl_games)	0.779	0.155	134.540	0.095	2.189	—
TrackVelocities
	FloatingPlatform (skrl)	0.930	—	—	0.447	—	0.049
	FloatingPlatform (rl_games)	0.679	—	—	0.388	—	0.044
	Kingfisher (skrl)	—	—	—	0.618	—	0.241
	Kingfisher (rl_games)	0.434	—	—	0.093	—	0.272
	Turtlebot2 (skrl)	0.768	—	—	0.152	—	0.107
	Turtlebot2 (rl_games)	0.783	—	—	0.025	—	0.100

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.

📊 Pre-trained models and performance metrics

You can download all the trained models from this link.

Simulation

Real-world

Turtlebot 2	Kingfisher	Floating platform