Deep Reinforcement Learning-based Three-dimensional Path Planning Algorithm for Tea Bud Picking

Abstract:

Picking path planning is the core decision-making link that determines the harvesting efficiency of automated tea bud picking robots. Aiming at the problems of difficult three-dimensional spatial modeling, high real-time response requirements and poor adaptability to dynamic node scales in automatic tea bud picking path planning, the tea bud picking task was modelled as a three-dimensional traveling salesman problem (3D-TSP), and a deep reinforcement learning-based 3D path planning algorithm was proposed for tea bud picking. A parametric 3D tea bud dataset with a zero-padding strategy was constructed, a dynamic masking mechanism was designed to ensure path validity, a long short-term memory (LSTM) module was introduced to dynamically fuse historical path information, a constrained multi-head self-attention mechanism was developed to improve decision-making accuracy, and the stability model training was done based on the Actor-Critic framework. Comparative experiments showed that the relative optimality gap between the proposed algorithm and the exact Concorde algorithm was controlled within 2% , and its computation time in the scenario with 80 tea buds was only 1.7% of that of Concorde. The algorithm outperformed classical heuristic algorithms and other deep reinforcement learning-based algorithms, maintained stable performance in real tea garden scenarios, and provided an efficient solution for tea picking robot path planning.