Target-conditioned flow-matching policy for citrus harvesting robot pre-grasp approach behavior learning

Fruit harvesting in natural orchards remains challenging because target fruits are distributed in cluttered and unstructured environments. In the pre-grasp approach stage of multi-fruit citrus harvesting, three issues are particularly critical: limited demonstration data, target ambiguity, and the need for stable and precise local approach motions.

To address these issues, this study proposes a Target-Conditioned Flow-Matching Policy (TCFM Policy), which integrates image observations, robot state history, and explicit target geometric conditions, uses a dual-branch visual representation to encode both global scene context and local end-effector details, and predicts future multi-step TCP trajectories through conditional flow matching. To reduce overfitting to global appearance under small-sample conditions, a target-oriented visual augmentation strategy is further introduced for the global branch during training.

A real-world dataset containing 160 valid demonstration episodes was collected on a UR5-based citrus harvesting platform using VR teleoperation. In 50 target-specified multi-fruit trials, the full model achieved a success rate of 76%, a target-picking error rate of 4%, and a picking-point offset rate of 20%.

A fairness-aligned comparison with a target-conditioned diffusion-policy baseline further shows that the proposed method achieves lower offline trajectory error and better online target-specified approach performance under the same training setting. Ablation results indicate that the ROI branch mainly improves final alignment, while the target-oriented augmentation mainly improves target consistency. These results indicate that explicit target conditioning, dual-branch visual encoding, and conditional flow matching jointly support accurate target selection and relatively stable pre-grasp approach execution in small-sample multi-fruit citrus scenes.