Safe Crane Handling with Collision-Free Path Planning using DDPG

• Rafaela Iovanovichi Machado FURG
• Matheus Machado Dos Santos FURG
• Silvia Silva Da Costa Botelho FURG

Resumo

Operating cranes can be an extremely challenging task due to the low visibility or blind spots in the work area. These problems with the operator visibility causes an alarming number of crane accidents. To improve crane operations, this study aims to develop a collision-free path planning model for crane manipulation, these simulators and training techniques offer a cost-effective and environmentally friendly way to conduct training exercises, without the need for real-life scenarios that could damage the equipment, harm workers or disrupt ongoing operations.In order to assist crane operations, this study aims to develop a collision-free path planning model for crane manipulation. Given that, this is a complex problem with challenging data acquisition, so the use of methodologies employed in supervised and unsupervised learning becomes impractical, as they require prior data for model training and testing.Hence, for this work’s development, it was decided to use reinforcement learning (RL) techniques. In RL algorithms the agent learns to take actions to maximize an accumulated reward over time, aiming to learn the best strategy of action in a given situation interacting with the environment to which it is inserted.During our evaluation of the Deep Deterministic Policy Gradient (DDPG) algorithm, we considered two distinct reward methods: an adapted reward and a proposed reward. Our analysis revealed that the performance of the proposed reward method outperformed that of the adapted method, displaying notably enhanced training efficacy. These findings underscore the considerable advantages of incorporating the suggested reward into our approach.

Palavras-chave: Reinforcement learning, crane, path