Mobile Robot Path Planning Using Deep reinforcement learning

摘要

This article provides an overview of a robotics project that explores path planning techniques using deep learning and reinforcement learning. The goal of the project is to enable map-less navigation for mobile robots. The project utilizes the Robot Operating System (ROS) and Gazebo simulation tools along with TurtleBot3 robots. Data is collected by moving the robot in different simulated environments and recording images, velocities, and positions. This data is used to train Vision Transformer (ViT) and convolutional neural network models to predict robot velocities based on visual input. The trained models are then utilized within a Soft Actor-Critic reinforcement learning algorithm to learn optimal navigation policies. Key challenges addressed include data collection, network architecture design, synchronization with the ROS framework, reward definition, and training adjustments to improve goal achievement. The project provides insights into deep learning techniques for robotic vision and control. It demonstrates the potential of combining imitation learning and reinforcement learning for map-less navigation. Further work could explore deploying the trained models to physical robots and enhancing generalizability across diverse environments. This research contributes to advancing vision-based autonomous robot navigation.