Low Compute End-to-End Robot Navigation from Depth Images Using Deep Reinforcement Learning
International Journal of Innovations in Science and Technology · 12.05.2026 · 447
DOI: 10.33411/ijist/1816 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Innovations in Science and Technology, online erschienen am 12.05.2026. Verfasst von Shahrukh Hussain und 2 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Effective and compute-efficient navigation is essential for mobile robots to operate autonomously in complex environments such as crowded places like airports and shopping malls. Sensors mounted on mobile robots play a vital role in decision-making by providing information about the surroundings. Among these sensors, depth cameras are a type of visual sensor that provides rich depth information of the surroundings, enabling robots to comprehend the 3-dimensional structure of the environment, assisting the robot in robust obstacle avoidance and navigation. In this paper, we aim to achieve autonomous navigation of a differential drive robot using only the depth information of the environment by employing a simple Convolutional Neural Network (CNN) architecture. CNN interprets the depth images captured by the depth camera and generates corresponding actions for the robot, while maintaining computational efficiency due to the limited computational resources of mobile robots. We employ Deep Reinforcement Learning (DRL) with curriculum learning paradigm to train our CNN in two Gazebo robotic simulation environments with increasing complexity. This approach increases the generalization of the model and enhances its adaptability in unobserved environments. The CNN learns to navigate autonomously using only depth information from the environment. The trained model is then evaluated by deploying it in an unseen simulation environment. Results show that the agent converged in 1,100 episodes in the primary environment. Furthermore, to demonstrate the model’s adaptability, it is deployed in a real-life laboratory environment where it achieved a 70% success rate after training for 1,000 episodes.
Zitieren
APA
Hussain, S., Wahid, A., & M., A. (2026). Low Compute End-to-End Robot Navigation from Depth Images Using Deep Reinforcement Learning. International Journal of Innovations in Science and Technology, 447. https://doi.org/10.33411/ijist/1816
Vancouver
Hussain S, Wahid A, M. A. Low Compute End-to-End Robot Navigation from Depth Images Using Deep Reinforcement Learning. International Journal of Innovations in Science and Technology. 2026:447. doi:10.33411/ijist/1816
BibTeX
@article{Hussain2026,
author = {Hussain, Shahrukh and Wahid, Abdul and M., Anayatullah},
title = {Low Compute End-to-End Robot Navigation from Depth Images Using Deep Reinforcement Learning},
journal = {International Journal of Innovations in Science and Technology},
year = {2026},
pages = {447},
doi = {10.33411/ijist/1816}
}
EUVIMED