Robotic arms are often required for high picking efficiency, path planning, and path smoothness for safflower harvesting in unstructured environments. This study aims to introduce a picking area clustering and a Target redirecting rapidly-exploring random tree (TR-RRT*) path planning. Firstly, a roller end-effector was designed with an effective picking area of 5 cm × 20 cm, according to the spatial distribution of safflower seed balls under natural conditions. A picking point clustering was also developed to divide the working range of the robotic arm into multiple sub-areas. Each cluster was designed to cover 1-3 picking points. The end-effector was used to harvest 1-3 safflowers in a single operation. A goal redirecting, a goal-direct and deflection expansion, and an artificial potential field (APF) tangential escape force strategy were integrated into the bidirectional RRT* framework to improve the path search efficiency and obstacle avoidance. Among them, the goal redirecting strategy continuously updated the target points during expansion to rapidly connect paths within opportunity windows. The goal-directed and deflection expansion allowed both search trees to extend from the nodes closest to the goal, thus maximizing progress toward the target while deflecting to avoid obstacles. The tangential escape force strategy introduced a tangential component into the conventional artificial potential field, enabling smooth sliding along obstacle boundaries when approaching them. As such, the module effectively avoided the path oscillation and target unreachability in the conventional APF. In path optimization, a combination of a greedy jump-point strategy, interpolated curvature optimization, and B-spline curve fitting was applied to smoothly adjust the curvature of path nodes and then generate high-order continuous trajectories. An obstacle avoidance reconstruction was also introduced to prevent the trajectory penetration through obstacles for the continuous, collision-free, and smooth motion of the robotic arm. A comparison was made on the TR-RRT* and seven algorithms in the dense, small-obstacle environments. The results demonstrated that the superior performance was achieved in the complex obstacle environments, where the path lengths (3892.05mm) were shortened by 9.16% and 8.22%, compared with the IBI-P-RRT*(4284.44 mm) and BI-RRT*(4240.45 mm), respectively; While the average planning time (0.30 s) was only 7.94% of that of RRT*(3.78 s) and 69.77% of BI-RRT*(0.43 s), respectively. Additionally, the steering angles of BI-RRT*(37.51°) and BI-APF-RRT*(39.61°) were 1.22 and 1.29 times larger than those of the improved algorithm (30.63°), indicating significant improvement of the path smoothness. Statistical quantitative experiments were conducted for the algorithm in different obstacle environments. The optimal performance was achieved in both path length and time consumption under various environments. Moreover, the TR-RRT* algorithm successfully planned paths in narrow passage obstacle environments. The tendency of conventional APF was to avoid the local oscillations during path planning in such scenarios. Physical harvesting tests further validated the effectiveness of the clustering and TR-RRT* algorithm. The robotic arm took an average of 3.64 s to move from the initial position to the first target point, with an average transfer time of 3.12 s between tasks. The average positional error relative to the robotic arm's workspace was less than 0.9%, indicating the stable and efficient performance of safflower harvesting.