Neural-Network-Based Key Pose Computation and Trajectory Planning for 7-DOF Manipulators in Individual Tomato Harvesting

In recent years, robotic tomato harvesting has attracted extensive attention. Many studies have achieved promising results in cluster harvesting of tomatoes. However, for large-sized tomatoes that require individual fruit harvesting, current automation systems exhibit high perception accuracy but still face significant challenges in manipulator pose computation and trajectory planning. To address this issue, this paper proposes a complete pose computation and trajectory planning framework for a 7-DOF manipulator dedicated to individual tomato harvesting. The framework includes a novel neural-network-based method, IK-Resnet-Jacobian, to solve the manipulator camera viewpoint pose, a Jacobian-based iterative method for computing the harvesting pose, and the RRT* algorithm for trajectory connections among the placing pose, camera viewpoint pose, and harvesting pose. Experimental results on a constructed dataset show that, compared with traditional inverse kinematics (IK) methods, the proposed IK-Resnet-Jacobian improves the success rate of solving camera viewpoint poses by 10.91%, while generating higher-quality poses with less trajectory redundancy. Tests conducted in a realistic physical environment simulating tomato harvesting further demonstrate that, compared with the previous pose computation and trajectory planning pipeline, the proposed framework increases the tomato cluster localization success rate by 37.14% and the harvesting success rate by 28.57%. These results confirm the superiority of the proposed pose computation and trajectory planning framework for robotic individual tomato harvesting.