ASABE Chen 19 Boston SV chen

Abstract

An emerging trend in agricultural field operations is the deployment of supervised autonomous teams of cooperating agricultural self-propelled machines. During such operations several machines (Agricultural Primary Units - APUs) perform the main field task (e.g., spraying, harvesting), and other machines (Field Service Units - FSUs) provide in-field logistics support, i.e., they transport working materials (chemicals, crop) between APUs and other units stationed outside the field. The same concept is being researched for robot-aided harvesting of specialty crops. For example, in strawberry harvesting, each picker (APU) harvests and fills up a tray; a mobile robot (FSU) travels to the picker and when their tray is full, it transports the tray to a field unloading station. Reactive scheduling policies for FSUs (go to a picker when tray becomes full) are not efficient, because FSUs may have to traverse large distances to reach pickers, thus introducing large waiting times.

This work investigates FSU predictive scheduling with perfect information and applies it to robot-aided strawberry harvesting. It assumes that the robots can predict exactly the time and the location of each picker’s next transport request (time when tray fills up). The reasoning behind predictive scheduling is that a robot can start going toward the location of a predicted request before the tray is filled, in an effort to reduce/eliminate waiting times due to traveling. Prediction is typically based on measurement of the rate that a tray (or a combine harvester’s tank) is filled. As more crop is harvested (and weight data collected) while the tray fills, predictions become more reliable. However, if prediction information for a transport request becomes available when the tray is almost full, predictive scheduling becomes reactive; having access to predictions earlier when a tray is still being filled is expected to increase performance.

This work simulates robot-aided strawberry harvesting – based on data gathered from a commercial operation in Salinas, CA – to investigate the effect of early/late access to perfect prediction information on the performance of a predictive robot scheduler. In the simulations, predictions of next tray transport requests become available at different tray fill-ratios FR (0% - empty to 100% 0- full) and the performance of dispatcher is respectively studied and analyzed. For the fields, yields and harvest crews modeled, it was found that the performance of the scheduler does not show explicit improvement when the fill-ratio is less than FR=0.9. The specific value of FR depends on the average time (s) it takes for robots to traverse the distance to pickers, and the remaining time to fill (1-FR) of the tray’s weight. It is expected that the longer it takes robots to reach pickers, the smaller FR should be (earlier prediction). To verify this, different robot travel speeds are simulated - leading to different robot arrival times at the predicted tray pick-up locations – and results will be presented.

Full text

Scheduling Performance of Harvest-aiding Crop-transport
Robots under Varying Earliness in access to Predictive
Transport Requests
Chen Peng1
Stavros G. Vougioukas2
UC Davis
1Department of Mechanical and Aerospace Engineering
2Department of Biological and Agricultural Engineering
2019 ASABE Annual International Meeting

2
Outline%
•Introduction
•Background
•Reactive vs. predictive scheduling;
•Earliness: Fill Ratio (FR);
•Main problems;
•Methodology
•System Architecture;
•Simulator;
•Processing predictive requests;
•Predictive scheduling;
•Results and discussion
•Effects of FR on scheduling performance;
•FRs on predictive scheduling;
•Analytical estimation of adequate FR;
•Summary and Future work

3
Background:%Manual Strawberry Harvesting Process
Field Setting
Collection station
Manual Harvesting
Tray delivery
•Pickers spend 15%-20% of
their time walking.
•This non-productive time
decreases efficiency.

4
Background:/ Crop-transport Robot
Field Setting
Transporting Robots
GOALS:
•Efficiency improvement.
•Labor reduction.
Collection station
Manual Picking

Motivation for efficient scheduling
5
Cost-effectiveness:
 Minimize robot to picker ratio.
 Robots are shared:
 Reduce waiting time of pickers.

6
First-come-rst-serve Reacve Dispatching
Mean and standard deviaon of
Motivation behind Predictive Dispatching)
Wait for robot to cover
distance from depot to
picker!!
Predicve
Scheduling
Settings:
•120 Furrows × 100m;
•15 Pickers;
•Robot velocity: 1.2 m/s;

7
Time (s)
Weight (g)
Fill Ratio
Tray capacity
Predicting Transport Requests: Fill
Ratio
Impractical to get the
Prediction result
Pickers State
Observaon Module
Data Cached
Linear
Regression
State
Recognition
Over FR?
Picker state
observation
GPS in SBAS
mode
Load Cell
sensors
Full tray time Full tray location

8
Main Problems
How does earliness of predicon aect scheduling performance?
How early are predicve requests enough? (FR=?)
Wait me vs. earliness of request predicon

Field Operations
Dispatching System
Methodology: System Architecture
Pickers Operaon Robots Operaon
Transport Request
Predicon
Predicve
Scheduling
Pickers State
Observaon Module
Transport Requests
Robots State
Observaon Module
Pickers states Robots states Dispatch
Commands

10
Simulator
Typical strawberry harvesting in Santa Maria:
•25 pickers in a crew;
•100 furrows *100m harvesting block;
•One active collection station;

11
Predictive scheduling
Problem formulation
•Inputs:
•Available transport
requests;
•Robot states;
•Objective: Minimize mean of
wait times;
Problem modeling
•Parallel machine scheduling problem:
•Temporal constraints;
•NP-Hard;
•Methodology:
•Branch and Bound Search
•Heurisc method Soluon Interpretaon
•Robot id;
•Serving Picker id;
•Dispatch locaon;
•Dispatch me;

12
Results and discussion
Very early availability of prediction
(when FR < 0.8) has no effect on
scheduling performance.
FR vs Waiting time

13
FR vs Early requests ratio
Early requests Late requests
Robot k
Request i
𝐸𝑎𝑟𝑙𝑦 𝑅𝑒𝑞𝑢𝑒𝑠𝑡𝑠 𝑅𝑎𝑡𝑖𝑜=𝐸𝑎𝑟𝑙𝑦 𝑅𝑒𝑞𝑢𝑒𝑠𝑡𝑠 𝑛𝑢𝑚𝑏𝑒𝑟
𝑇𝑜𝑡𝑎𝑙 𝑅𝑒𝑞𝑢𝑒𝑠𝑡𝑠𝑛𝑢𝑚𝑏𝑒𝑟
∆𝑡𝑖
𝑓
∆𝑡𝑖
𝑓

14
Efficiency vs FR at different robot speeds
•Smaller robot velocity, earlier
requests needed;
•“Turning point” shift to the right;

15
Required FR estimation
An estimate for :
•Field size:
•Robot velocity ()
Robot velocity (m/s)
0.8 0.75 0.68
1.0 0.80 0.75
1.2 0.82 0.79
1.5 0.85 0.83
1.7 0.87 0.85
2.0 0.90 0.87

16
Conclusions
•When the robot/picker ratio is near 1:3, the mean of non-productive time is reduced by
approximately 60s compared with manual harvesting, 84.99s.
•For the specific commercial field operation and harvesting scenario, the FR starts affecting the
performance of predictive scheduling when a certain value is reached (~ 0.88).
•A lower bound of FR start-impact value can be estimated given a specific harvesting
configuration including harvesting parameters of pickers, field dimension, and robot velocity.
Future Work
× Evaluate scheduling performance with the consideration of stochasticity of
transport request prediction
× Field experiments

17
Acknowledgements
This work is part of the National Robotics Initiative project titled "FRAIL-bots:
Fragile cRop hArvest-aIding mobiLe robots", that is funded by NIFA-USDA