Título: TD3-Based DRL for Skid-Steering Robot Navigation in Multi -Scale Environments
Autores: Carlos J. F. S. Júnior, Carlos Eduardo da Silva Cerqueira, André G. S. Conceição & Tiago Trindade Ribeiro
Resumo: This work explores the use of Deep Reinforcement Learning (DRL) for autonomous navigation of a skid-steering ground robot. The Twin-Delayed Deep Deterministic Policy Gradient (TD3) algorithm is employed to train a control policy that guides a LiDAR-equipped robot to reach relative targets while avoiding obstacles. The approach is evaluated in environments with varying dimensionality, demonstrating convergence and the ability to generalize navigation strategies across different scenarios. However, training in large-scale environments presented significant challenges, particularly in achieving stable convergence. The results indicate a strong dependence on the relative dimensions of the environment, with policy performance tending to degrade in overly sparse or disproportionate spatial configurations.
Palavras-chave: Skid-steering robot; Autonomous navigation; Deep reinforcement learning; TD3 algorithm; Obstacle avoidance.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1188151
Artigo em PDF: CBIC_2025_paper1188151.pdf
Arquivo BibTeX:
CBIC_2025_1188151.bib