Nianzu Dai ,  Jiaming Fang ,  Jin Yuan ,  Xuemei Liu

Abstract

 Robotic tomato harvesting in dense greenhouse clusters remains difficult because local visibility, accessibility, and contact safety change during observation and execution. Existing harvesting pipelines often separate perception from execution, which can lead to redundant views or failed harvesting attempts. This study proposes LARIS, an execution-oriented framework that integrates local-state reasoning with feasibility-guided observation–harvesting decision-making. LARIS formulates an integrated local execution-state representation to describe target geometry, neighboring local-structure evidence, free-space evidence, and residual uncertainty. Based on this state, it switches between supplementary observation and harvesting according to execution sufficiency rather than visual completeness alone, and adapts approach, contact, and retraction margins to residual uncertainty. In greenhouse experiments, LARIS achieved a harvesting success rate of 79.6% (133/167; 95% CI: 72.9–85.0%), with 1.60 ± 1.33 supplementary observations and a cycle time of 8.71 ± 1.33 s per attempt. In indoor reconstructed-cluster experiments, LARIS achieved a success rate of 84.1% (127/151; 95% CI: 77.4–89.1%). These results indicate that LARIS can improve harvesting reliability and efficiency under the tested greenhouse conditions with dense tomato clusters.

Paper Linkage:https://doi.org/10.1016/j.compag.2026.112312