Deep Reinforcement Learning-Based Coverage Path Planning: A Comparative Analysis of the Impact of Curriculum Learning on Environments with Dynamic and Static Obstacles Derin Peki stirmeli grenme Tabanli Kapsama Yol Planlamasi: M fredat ?grenmesinin Dinamik ve Statik Engelli Ortamlardaki Etkisinin Kar sila stirmali Analizi

Kuscu T., İNNER A. B.

33rd IEEE Conference on Signal Processing and Communications Applications, SIU 2025, İstanbul, Türkiye, 25 - 28 Haziran 2025, (Tam Metin Bildiri) identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/siu66497.2025.11111938
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: Coverage Path Planning, Curriculum Learning, Deep Reinforcement Learning, Dynamic Environments
  • Kocaeli Üniversitesi Adresli: Evet

Özet

Coverage Path Planning (CPP) is crucial for autonomous robots to systematically cover an area. Achieving optimal coverage in dynamic environments with obstacles presents challenges like position uncertainty and adaptation. This study explores curriculum learning in a deep reinforcement learning-based CPP approach. Using Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), standard single-stage training and curriculum learning were compared in grid-based environments with static and dynamic obstacles. The proposed curriculum gradually increases difficulty by adjusting obstacle count, speed, and coverage thresholds. The results indicate that curriculum learning enhances performance in both algorithms, with improvements of 8% in average coverage and 30% in full coverage for DQN, and up to 4% and 18%, respectively, for PPO. Additionally, the curriculum-trained agent had shorter path lengths despite higher revisit rates, suggesting it prioritizes covering more area. This study underscores the effectiveness of curriculum-based deep reinforcement learning for CPP, especially in complex environments.