Answers are grounded in this paper and cite their sources.
Experimental evaluation demonstrates real-time branch segmentation and high harvesting accuracy in robotic litchi harvesting, suggesting practical viability for automated agricultural systems.
Accurate picking point localisation is a fundamental challenge in robotic harvesting. For litchi, this requires precise identification of the main fruit-bearing branch (MFBB). Current methods, however, rely on implicit structural association, failing to learn the explicit spatial correlation between fruits and branches, which leads to frequent misidentification. To address this, LitchiInst is proposed, a real-time instance segmentation framework that explicitly models this fruit-branch dependency. The core of LitchiInst is a Structure Association Guidance (SAG) mechanism, which employs SAG queries and a dedicated SAG loss to enforce spatial correlation and enable structure-aware discrimination. The framework is further enhanced by an Enhanced Feature Pyramid Network (EFPN) to capture fine-grained MFBB features and a Mask-to-3D head to convert segmentation masks into stable 3D picking points for robotic execution. LitchiInst improves MFBB average precision by 31.06% relative to the baseline while maintaining real-time processing. In real-robot experiments across three testing settings, LitchiInst achieved a 92.0% MFBB detection success rate. These results further demonstrate its practical potential as a visual guidance module for automated litchi harvesting.
No takes yet. Share an insight, caveat, or question.
Qian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: