This paper presents a cloud robotics architecture for greenhouse strawberry harvesting, designed as a two-phase workflow that decouples perception from actuation. This study aims to demonstrate the feasibility and scalability of cloud-based perception offloading in structured agricultural environments.
In Phase 1, the robot performs single-pass data collection, capturing red, blue, green, and depth images and transmitting them to a cloud server for processing. In Phase 2, harvesting is executed based on cloud-computed localization results. Three models are developed: error propagation for localization accuracy, cloud latency for permissible delays and probabilistic harvest success linking detection, localization and gripping reliability.
Experiments confirmed localization repeatability of 0.6 mm root-mean-square within the 11 mm tolerance of the cutter–gripper, harvesting success rates of 97.9% in open/cluttered arrangements and 81.7% in overlapping clusters and average cycle times of 1.6 s per plant for data collection and 7.8 s per fruit for harvesting. Cloud latency tests yielded delays from 1.6 to 13.8 s depending on network condition, validating the latency model. Cost analysis showed that lightweight robots using low-cost processors (e.g. Raspberry Pi 4) can leverage cloud resources, avoiding expensive onboard graphical processing units and enabling scalable multi-robot deployment.
To the best of the authors’ knowledge, this study demonstrates the first integration of single-pass data collection with cloud-based perception offloading for strawberry harvesting. The findings provide a generalizable framework for cloud-integrated service robots, with applicability to other greenhouse crops and broader service robotics domains.
