<![CDATA[This research addresses the issue of automatic obstacle avoidance during navigation by autonomous agents. Designing a traditional, programmed automatic obstacle avoidance system would be difficult and expensive. Therefore, a neural network-based approach is proposed, known as end-to-end deep imitation learning, where the approach is data-driven and thus relatively easier and more cost-effective compared to traditional methods. The research also proposes the architecture of a convolutional neural network design and image processing techniques for effective and efficient machine learning training. Testing is conducted on a path with randomly placed obstacles in the Webots simulator. Gradual performance evaluations demonstrate that the proposed architecture successfully trains autonomous agents to maneuver when encountering dynamic obstacles with a relatively small training dataset.]]>
Copyrights © 2023
