Abstract: Hilly and mountainous terrains are fragmented with significant slope variations. In such environments, paths planned by conventional path planning algorithms are prone to frequent turning, high energy consumption, and poor safety margins, making it difficult to meet the practical demands of autonomous agricultural machinery operations. To address the path planning problem in hilly and mountainous agricultural operation areas and to enhance the operational efficiency and energy efficiency of intelligent agricultural machinery, this study proposed a path planning method for agricultural machinery based on an optimized Theta* algorithm. First, using airborne LiDAR point cloud data, a progressive morphological filter was applied to remove non-ground objects such as vegetation and buildings, followed by a strip curve fitting method to extract ground points. Based on the filtered ground points, a digital elevation model (DEM) of the operational area was constructed to provide accurate three-dimensional terrain information for subsequent path planning. Second, a multi-objective cost function was established, which incorporated three-dimensional Euclidean distance, normalized energy consumption covering both straight-line travel and spot-turn steering, and a safety cost derived from an obstacle distance field. The energy consumption was normalized to the equivalent level-ground travel distance, unifying the dimensions and magnitudes of different cost components for effective weight tuning. To balance search efficiency and the local optimum problem inherent in Theta*, an adaptive weight factor was introduced. This factor enabled the exploration strategy to gradually shift from depth-first to breadth-first as the search progressed, thereby ensuring rapid convergence in the early stage while preserving path quality in the later stage. Third, the line-of-sight detection mechanism was improved by integrating interval sampling-based slope verification with accurate three-dimensional cost calculation. Along each potential line-of-sight connection, equally spaced sampling points were generated, at which both collision detection and slope angle verification against the vehicle's maximum uphill and downhill climbing limits were performed. This ensured that direct parent–child node connections were not only collision-free but also satisfied the slope traversability constraints. Furthermore, a backtracking mechanism was designed to strictly enforce the slope passability constraints of tracked agricultural machinery during node expansion. Simulation experiments were conducted using the open-source DCPP dataset, which represents a terraced orchard scene in hilly terrain with significant elevation variations. Ablation experiments and comparative experiments were carried out respectively. Ablation results showed that individually varying each of the four cost weights could drive the corresponding path metrics to change in the expected directions, while an excessive increase in the heuristic function weight led to a significant decline in overall path quality. In the comparative experiments, the proposed optimized Theta* algorithm was benchmarked against basic Theta*, basic A*, optimized A*, the improved gray wolf optimizer, and an improved Theta* algorithm. The results demonstrated that the optimized Theta* consistently found paths with superior comprehensive performance across multiple metrics. Compared with basic A*, the optimized Theta* achieved a 4.20% reduction in path length, 35.34% reduction in total energy consumption, 43.65% reduction in average safety cost, and 93.93% reduction in turning frequency. These findings indicated that the proposed method could effectively enhance the smoothness, economy, and safety of agricultural machinery operation paths in hilly and mountainous areas, thereby providing valuable technical support for the autonomous navigation of intelligent agricultural machinery in complex three-dimensional terrains. In contrast to basic Theta*, the proposed method did not evaluate visible node connections solely in a planar grid. It coupled terrain-aware visibility verification with the multi-objective cost model, so that each candidate connection was examined in terms of collision avoidance, slope traversability, three-dimensional distance, and energy consumption. The adaptive search strategy and backtracking mechanism further reduced the risk that an apparently short connection violated vehicle constraints. Consequently, the resulting path reflected not only geometric efficiency but also the operating characteristics of tracked agricultural machinery on uneven terrain.