Modelo para Planejamento de Caminho Livre de Colisão para Manejo de Guindaste utilizando DDPG

Abstract

To aid the crane operators this work proposes the integration of Reinforcement Learning (RL), using Deep Determinism Police Gradient (DDPG) based on the concepts used in autonomous cars, to generate collision-free paths. The proposed work uses a digital twin, which is a virtual replica of the physical system to perform agent training, the reward function used in this study allow the network to learn from its mistakes and optimize its behavior to reach the target position efficiently. Punishments and rewards are given to the agent in the form of positive or negative scores, and when the laser does not identify any object, the agent will not receive any score. The success of the network’s performance on the Gazebo simulator environment demonstrates its potential to solve complex problems in the real world. With further development and finetuning, this approach could have practical applications in industries such as manufacturing and logistics. Additionally, the use of virtual environments to test and validate the network’s performance can lead to significant cost savings and reduced risk compared to physical testing.