Reward Shaping for Mobile Robot Navigation Based on Deep Reinforcement Learning.
Published in Researchgate, 2024
Abstract: In recent years, various robot navigation method has been proposed, including the use of SLAM (Simultaneous Localization and Mapping), various path planning approaches such as A* or Djikstra, and obstacle avoidance methods such as an Artificial Potential Field. A navigation method is important for mobile robots to navigate autonomously. However, there are still challenges to including various constraints and diverse sensor data in the navigation process, particularly if no maps of the environment are available. The characteristic of an uncertain environment led to the development of a navigation method with Deep Reinforcement Learning (DRL). However, to solve the navigation problem, the DRL algorithm needs a reward shaping that generates a suitable reward function with limited observation data. In this paper, we proposed reward shaping for Deep Q Network (DQN), which is one of the methods of the DRL algorithm to perform navigation for a mobile robot. The goal of the algorithm is to achieve a specific position with optimum accuracy on a 2-dimensional space, making use of positioning data only. We compared a positive and negative reward function to train the DQN. We also compare two types of robot observation. We have successfully achieved the best performance of positioning with 87.2% accuracy.
Recommended citation: MBadriawan, Yusuf & Cahyadi, Adha & Rizqi, Ahmad Ataka Awwalur & Setiawan, Noor. (2023). "Reward Shaping for Mobile Robot Navigation Based on Deep Reinforcement Learning." Conference 1. 1(1).
Download Paper