Robotic harvesting systems for greenhouse agriculture are often designed as monolithic platforms that tightly couple perception, planning and actuation, requiring repeated real-time sensing and high-performance onboard computing. This study aims to develop and validate a cloud-coordinated harvesting architecture that separates perception and harvesting across independent robotic platforms, enabling single-pass plant-level data collection, reduced onboard computation and scalable multirobot operation in structured greenhouse environments.
A cloud-coordinated dual-robot harvesting system was developed consisting of a data collector robot for perception, a harvesting robot for execution and a cloud server for perception processing and coordination. Analytical models characterize perception-side and execution-side errors, cloud latency constraints and a system-level harvesting feasibility condition. Experimental validation was conducted in a controlled greenhouse environment, including world-frame localization accuracy evaluation, cloud latency testing, execution-side positioning accuracy assessment and harvesting trials under varying fruit configurations.
Results show a mean world-frame localization error of 2.6 mm and a mean execution-side positioning error of 3.0 mm. Cloud-based perception offloading remained robust under network degradation, while lightweight cloud-to-robot data retrieval enabled responsive harvesting execution. Harvesting success rates ranged from 86.5% to 98.5% with an average harvesting cycle time of 8.8 s per fruit.
This study introduces a cloud-coordinated dual-robot harvesting architecture with a cross-robot error-propagation framework linking perception accuracy, execution uncertainty and end-effector tolerance, enabling reusable plant-level data and reduced onboard computational requirements.
