RRT-Guided and Spatial Attention-Based TD3 Deep Reinforcement Learning Method for Safflower Filament Harvesting Path...

Path planning for safflower filament harvesting is a critical component for ensuring the efficient operation of robotic harvesting systems. This paper proposes a path planning algorithm named SA-TD3-RRT, which integrates the Rapidly-exploring Random Tree (RRT), Twin Delayed Deep Deterministic Policy Gradient (TD3), and a spatial attention mechanism to achieve hierarchical planning of safflower filament harvesting paths. In the RRT component, techniques including adaptive step size, goal-biased sampling, and Douglas-Peucker path simplification are introduced. Specifically, the adaptive step size strategy converts the fixed step size into a dynamically adjustable extension step, effectively improving both the search real-time nature and path quality of the algorithm. The purely random sampling is replaced with goal-biased sampling, guiding the search tree growth toward the goal while suppressing redundant nodes, thereby reducing path planning time. Post-processing techniques such as Douglas-Peucker are employed to significantly reduce path complexity and improve path quality. In the reinforcement learning component, a spatial attention module is added, enabling dynamic perception of obstacle distribution. A spatial awareness term is introduced into the reward function to obtain more favorable spatial attention rewards. Additionally, the smoothed RRT path is injected into the experience replay buffer as expert demonstrations, realizing policy guidance based on imitation learning. Simulation results show that under favorable operating conditions, the proposed algorithm reduces the planned path length by 85.64%, 82.69%, and 92.55% compared to TD3, DDPG, and PPO algorithms, respectively, and shortens the path planning time by 17.68% relative to the original TD3 algorithm. In a three-dimensional environment, the planned path length is reduced by 8.29% compared to TD3, and the planning time is decreased by 81.24%, reaching 4.72 s. Although the current results are based on rigorous evaluation and demonstration in a high-fidelity simulation environment that abstracts from unstructured planting conditions, the proposed SA-TD3-RRT algorithm consistently achieved significant performance gains over the baseline. This simulation-based feasibility study provides a solid algorithmic foundation and key parameter benchmarks for future deployment on physical harvesting platforms, effectively bridging the gap between algorithmic innovation and practical agricultural robotics.