Data-driven trajectory optimization in robotic fruit harvesting via deep learning-based perception, gripper configuration, and fruit morphometrics

Abstract Conventional trajectory planning methods for robotic fruit harvesting mainly rely on static geometric heuristics and often overlook critical sensory and task-specific variables such as fruit morphology and end-effector compatibility. These limitations make traditional approaches less effective in real-world agricultural settings, where conditions are unpredictable and fruits require careful, adaptive handling. Moreover, most existing studies do not incorporate a Convolutional Neural Network (CNN) to detect confidence in the planning process, often treating perception and motion planning as isolated components rather than a unified system. To overcome these challenges, this study proposes a data-driven approach to trajectory optimization that integrates visual perception based on CNN confidence levels, gripper type with different actuation technologies, and fruit orientation, parameters that significantly influence harvesting efficiency. Two multivariate regression models were developed, one specifically for firm fruits such as oranges and the other for soft fruits such as strawberries. The models predict trajectory length using three input variables: CNN detection confidence, actuator type, which includes three-finger and two-finger grippers, and fruit orientation angles ranging from 50°–130°. The non-linear influence of orientation is captured through polynomial terms. A total of 46 experimental trials were conducted for each fruit type using a robotic platform under controlled conditions. The regression outputs revealed that CNN confidence had a strong influence on trajectory length reduction, while orientation had a more severe impact on strawberries due to their delicate structure. In comparison to baseline trajectories, the optimized A* planner, guided by regression coefficients, curtailed trajectory lengths by 11% for strawberries and 14% for oranges. Moreover, the positional accuracy incre ased by 15% and 12%, respectively. The higher predictive accuracy was attained by the models (R 2 = 0.89 and 0.82; RMSE = 3.2 cm and 4.7 cm for strawberries and oranges, respectively). These results demonstrate that heuristic planning, combined with statistical modeling, enhances motion reliability and spatial efficiency in autonomous fruit picking.