<p>Accurate acquisition of grapevine morphological information is a critical prerequisite for autonomous robotic pruning, yet remains challenging due to complex geometry, severe self-occlusion, and partial observability. Existing grapevine perception systems largely rely on passive sensing strategies and fixed acquisition trajectories, which often result in incomplete or uncertain reconstructions. This paper presents an information-theoretic active perception framework for robotic grapevine morphological reconstruction, in which camera viewpoints are selected to maximise expected information gain with respect to a probabilistic belief over vine structure. Grapevine morphology is modelled as a hybrid discrete–continuous state capturing both the existence of structural elements and their geometric attributes. An RGB-D perception pipeline is developed to detect vine components, infer topology, and update belief states incrementally as new observations are acquired. A tractable decomposition of information gain enables efficient evaluation of candidate viewpoints, while explicit soft-revisit penalties are introduced to mitigate degenerate action repetition and oscillatory behaviour commonly observed in myopic planners. The proposed approach is implemented on an eye-in-hand robotic sensing platform and evaluated on real grapevine specimens. Experimental results demonstrate that active viewpoint selection systematically improves structural completeness and geometric accuracy by resolving occlusions and reducing uncertainty compared to passive sensing strategies. The study further highlights practical considerations in information-theoretic planning for complex agricultural environments. Overall, the results demonstrate the effectiveness of active perception for grapevine information acquisition and provide a principled foundation for perception-driven robotic pruning systems.</p>