Comparative Evaluation of the Generalization Performance of Reinforcement Learning Algorithms for Agricultural Mobile Robot Navigation using Curriculum Learning
2026 11th International Conference on Computer Science and Engineering (UBMK), İstanbul, Türkiye, 16 - 20 Eylül 2026, ss.1630-1635, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ubmk71847.2026.11745196
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1630-1635
- Yozgat Bozok Üniversitesi Adresli: Evet
Özet
This study presents a systematic comparison of Reinforcement Learning (RL)-based approaches and a classical heuristic controller for the navigation of a differential-drive mobile robot operating in agricultural environments. Six Gazebo simulation scenarios (S0–S5) of increasing difficulty were designed; PPO, SAC, and TD3 algorithms were trained using a curriculum learning strategy on scenarios S0–S3 and subsequently evaluated over 10 independent runs on unseen scenarios S4 and S5. Experimental results show that TD3 achieved a 100% success rate in scenario S4, but declined to 10% in the more challenging S5 scenario involving dual obstacles and irregular plant rows. SAC demonstrated a more balanced generalisation performance with 80% success in S4 and 50% in S5, suggesting that entropy regularisation enhances robustness under complex obstacle configurations. PPO failed to reach sufficient learning performance within the given training budget and physics-based simulation constraints, and was therefore excluded from the final comparative analysis. The classical heuristic controller failed in all S4 and S5 trials, recording an average of 2.4 collisions in S4, 9.3 collisions in S5, and timeout in every run. The findings demonstrate that off-policy RL algorithms trained with curriculum learning can generalise to previously unseen obstacle configurations in agricultural environments; however, algorithm selection introduces critical performance differences depending on scenario complexity.