Muslikhin, Suprapto, Dwi Sudarno Putra, Anggun Winursito, Febrianto Amri Ristadi, Iwan Nofi Yono Putro, Ming-Shyan Wang
Trajectory planning on an eye-in-hand robotic manipulator with a cluttered heterogeneous environment is crucial to grasping success. This study proposes a hybrid model that combines deep learning-based environment detection with reinforcement learning (RL) to determine the optimal route. You Only Look Once (YOLO) was proposed in this system to incorporate deep reinforcement learning by optimizing the environment assessment. Merging YOLO with deep reinforcement learning optimizes environmental assessment by inserting a reward subprocess. YOLO assesses the environment for detection and recognition; after identifying targets or obstacles, RL can then reward or punish the agent to achieve the highest score. The entire process is encapsulated within the environmental cycle, which incorporates a state reward that is analyzed by deep learning. On the agent side, it is also driven by reinforcement learning. In this order, each agent and environment reach their optimal values, resulting in a 11-17% decrease in episode convergence time. © 2026, Department of Agribusiness, Universitas Muhammadiyah Yogyakarta. All rights reserved.
Electronics and Informatics Engineering Education Department, Universitas Negeri Yogyakarta, Indonesia; Automotive Engineering Education Department, Universitas Negeri Padang, Indonesia; Mechanical Engineering Education Department, Universitas Negeri Yogyakarta, Indonesia; National Research and Innovation Agency, Indonesia; Electrical Engineering Department, Southern Taiwan University of Science and Technology, Taiwan
