park_n_pi

MSc Thesis Project - University of Pannonia

Project Overview

This project focuses on developing a reinforcement learning (RL)-based goal navigation system for an autonomous car platform using a Raspberry Pi 5 and the Raspbot V2 chassis.
The core objective is to enable the robot to reach designated goal poses efficiently and reliably using deep reinforcement learning techniques, serving as a foundation for future autonomous parking maneuvers.

All experiments and evaluations presented in this project were conducted within a high-fidelity Webots simulation environment, integrated with ROS 2 Jazzy and Gymnasium.

Real-world deployment onto the Raspberry Pi 5 is planned for future extensions.

Features

• PPO-based reinforcement learning policy (Stable-Baselines3)
• ROS 2 Jazzy full simulation integration
• Custom vectorized Webots environments for faster training
• Curriculum learning for progressive goal-reaching difficulty
• Dockerized setup for easy reproducibility
• TensorBoard support for training monitoring
• Modular architecture supporting future real-world deployment

Project Structure

park_n_pi/
├── docker/
│   └── Dockerfile
├── ros_park_n_pi/
│   ├── src/
│   │   ├── raspbot_rl_env/
│   │   ├── raspbot_rl_interface/
│   │   ├── webots_py_ros2_driver/
│   │   └── raspbotv2_bringup/
├── scripts/
│   ├── run_wo_vecenv.py
│   ├── eval_wo_norm.py
│   └── (etc.)
├── logs/
│   ├── checkpoints/
│   ├── tensorboard/
├── resource/
├── README.md
├── .gitignore
└── LICENSE

Getting Started

First, build the Docker image:

docker build -f docker/Dockerfile -t park_n_pi_docker .

Then start the container using Docker Compose:

cd docker
docker compose up -d

Enter the container:

docker exec -it park_n_pi_dev bash

Inside the container, launch the Webots simulation and ROS2 nodes:

ros2 launch raspbotv2_bringup rl_training.launch.py num_robots:=8

Open a new terminal and connect again to the container:

docker exec -it park_n_pi_dev bash
cd src/park_n_pi/ros_park_n_pi/src/raspbot_rl_env/scripts/
python3 run_wo_vecenv.py --num_envs 8 --total_timesteps 1_500_000 --tb_log_name RunName

This starts the reinforcement learning training loop!

Usage

• Simulation and ROS2 nodes start automatically.

• Training runs with Stable-Baselines3 agents.

• Logs and models are saved under /home/dev_ws/src/park_n_pi/logs/

Stopping everything:

docker compose down

Rebuilding if needed:

docker compose down
docker compose build --no-cache
docker compose up -d

Technologies Used

• Docker

• ROS2 Jazzy

• Webots R2025a

• Stable-Baselines3

• Python 3.10+

• Gymnasium

License

This project is licensed under the MIT License. See the LICENSE file for details.