UC Davis

Successfully automating the harvest of fresh-market apples requires that robots match or exceed the apple picking throughput (FPT) of humans, at 0.83 fruits/s, while achieving a high fruit picking efficiency (FPE, percent of harvested fruits). To this end, robotic harvesters are built with multiple arms, though this introduces challenges. Fruit distributions are non-uniform, requiring workload balancing between arms. This can be accomplished through strategies such as flexible arm workspaces, having the harvester in constant motion, and choosing the right scheduling algorithm.

To address the need for workload balancing, we simulated a constantly moving harvester with multiple 3-degree of freedom arms sharing column workspaces. The arm’s movements were constrained by the column’s frame and software-defined vertical limits which prevented collisions between the arms. Initially, the harvester scheduled 3.5 m sections of orchard rows. It knew the location of all fruits, so could optimize its vertical arm limits, choose a “best,” fixed harvester speed, and compute the schedule for that orchard row segment. Two scheduling algorithms were evaluated, First Come First Served (FCFS) and a Mixed Integer Linear Programming formulation based on Goal Programming (GP). Later, harvesting was extended to the whole orchard rows using a new, semi-dynamic strategy, the Sliding Planning Window (SPW). The orchard row was solved as a series of individual, short, sequential, and overlapping planning windows which maximized the cumulative Orchard Row-FPT (OR-FPT) and OR-FPE.

Simulation experiments with real fruit distribution data validated our approach, resulting in combined throughput and efficiency gains for multiple harvester configurations. Increasing the number of arms increased the combined FPT and FPE. Using FCFS and the best harvester speed for nine arms, partitioning the columns to make rows of arms containing the samenumber of fruits resulted in a mean of 1.374 fruits/s compared to 1.049 fruits/s when the rows of arms were all the same height. Both achieved at least 95% FPE. Scheduling with GP improved the mean FPT to 2.0 fruits/s. When harvesting whole orchard rows, SPW always achieved at least 95% OR-FPE, however the amount of overlap between planning windows affected OR-FPT. No overlap resulted in a mean OR-FPT of 1.0 fruits/s. Overlapping half the previous planning window achieved 1.86 fruits/s. Further increasing the overlap decreased the OR-FPT. These results show that workload balancing could be used to increase the throughput of multi-armed harvesters and introduces a way to do this semi-dynamically.