Online prediction of tray-transport request time using mechanistic grey box models for improved scheduling of robotic strawberry harvest-aids

This paper addresses the online prediction of the time when a strawberry picker will fill up the carton tray (s)he is currently harvesting with. This time is important for predictive scheduling of tray-transporting harvest-aid robots that transport trays instead of the pickers themselves, with the goal to increase efficiency. To make such predictions, two mechanistic grey models (GM1 and GM2) were developed. The models incorporate knowledge about the process of filling a tray with harvested fruits by estimating relevant process parameters. During harvesting, both models compute Bayesian updates of the a priori distributions of the picking time and the tray's mass when it is empty and full, with data from each full tray. Moreover, both models rely on real-time measurements of the mass of harvested crop to predict when the tray capacity will be reached. GM1 predicts the next tray transport request (NTTR) time by adding the a priori mean picking time obtained from a previously developed predictive linear mixed model (LMM), and the beginning time of a cycle, whereas, GM2 fits a simple linear regression (SLR) to the load cell data in real time. In order to evaluate the performances of the two models, a set of load cell output time series was obtained from an instrumented picking cart during manual strawberry harvesting. Except for one tray (out of 58), both GM1 and GM2 were successful at generating NTTR predictions. However, GM2 predictions adapted to the occurrence of non-picking activities that manifested as plateaus in the mass data. Also, when a tray was not filled completely, GM1 could not predict accurately the NTTR time of that tray neither the NTTR of the next tray, because it could not detect the starting time for filling the next tray. GM2 missed the starting time also, but was able to provide accurate predictions, albeit a bit later in the picking cycle. However, delays that do not extend beyond 70% of the picking time are not significant. For these reasons, it seems that GM2 should be preferred over GM1. Additionally, since both GM algorithms utilise the a priori statistics of picking time and the full tray's mass, it is advisable to have pickers deliver a couple of trays themselves in the beginning of their shift, so that the process parameters are “learned” with Bayesian updating.