Slip‐Aware Coverage Path Planning Using a Turning Feasibility Table for an Orchard Robot Sprayer

ABSTRACT This study aimed to develop a coverage path planning (CPP) method for an autonomous vehicle equipped with a towed sprayer for orchards. In orchards, CPP can be formulated as a vehicle routing problem (VRP) in which work must be completed without duplication while appropriately supplying materials and fuel, as the work paths are already determined. Orchards present severe constraints, such as narrow headlands and the requirement for forward‐only turns due to the presence of a towed implement, resulting in many combinations of paths between which turning is infeasible. To achieve the objective under these conditions, there were two main challenges: i) turning feasibility is expected to depend on slippage, and ii) it is difficult to obtain a valid solution that does not cause dead ends when there are many combinations of paths that cannot be traveled. These challenges represent practical issues that were not considered in previous studies on route optimization. To solve challenge i), turning paths assuming forward turning were generated, and control simulations were performed on the paths to determine whether turning was possible. To account for slippage, a kinematic model was introduced that considers the slip angle of the vehicle's front wheels (steering wheels), assuming a normal distribution, based on previously acquired steady‐state circular turning data. Two types of traveling tests were conducted in an orchard with towing weights of 300 kg (0 L of liquid) and 700 kg (400 L of liquid) to verify the model's accuracy. As a result, the model error for the maximum deviation was improved compared with a model that does not consider slippage, with improvements of up to 65% for a towing weight of 300 kg and up to 51% for a towing weight of 700 kg. Next, to solve challenge ii), a turning feasibility table was created based on the simulation results proposed in this study, and an algorithm was devised to generate initial solutions for paths that reached without dead ends. As a result, while the random method could find initial solutions in 4% of cases, the proposed method achieved 13%, a threefold improvement in efficiency. Furthermore, the proposed method's evaluation value of initial solutions was good in 82% of cases compared to the random method, verifying that it is possible to search for effective initial solutions while ensuring diversity. Finally, CPP was performed using simulated annealing with the 2‐opt method based on the explored initial solutions. The optimization function explicitly considers recovery time in case of failure, in addition to the non‐working time commonly considered in previous studies. The formulation reduced overall lost time by 19%, and it was verified that optimization could reduce time by up to 23%. Through this series of methodologies, a CPP for orchards that is applicable in real‐world settings was achieved.