Online Point Coverage Path Planning for Prior-Free Robotic Weeding Using Deep Reinforcement Learning

作者：Zihan Liu, Jiale Wang, Chuyu Liu, Zehao Li, Hao Jiang, Yike Ma, Yucheng Zhang, Zhaoqi Wang · 年份：2025 · DOI：10.36227/techrxiv.175338547.70910319/v1 · 研究领域：Robotic Path Planning Algorithms、Robotics and Sensor-Based Localization、Control and Dynamics of Mobile Robots

The growing labor shortage in agriculture threatens global food security and accelerates the demand for advanced automation. Robotic autonomy offers an efficient alternative to labor-intensive tasks such as pasture weeding, which is essential for sustaining forage quality and livestock productivity. However, existing robotic weeding solutions often rely on strong environmental priors, such as known map layouts and obstacle-free terrain, which rarely hold in real-world fields. In this paper, we formalize the Online Point Coverage Path Planning (OPCPP) problem, where curvature-constrained robots must efficiently explore, avoid obstacles, and remove sparsely distributed weeds in unknown and unstructured environments under partial observability. To address the multi-objective and high-uncertainty nature of OPCPP, we propose an end-to-end deep reinforcement learning framework featuring three key components: (1) a multiscale diffusion feature representation to enhance sparse but behaviorally critical semantic cues; (2) a differential diffusion reward structure to provide dense and stable learning signals for multi-objective navigation; and (3) a dynamic curriculum training strategy that progressively scales task difficulty and reward priorities. Extensive simulations across diverse field scenarios show that our method outperforms both rule-based and state-of-the-art learning-based baselines, achieving up to 30% improvements in path efficiency and safety. Furthermore, real-world dep...