robotic apple harvesting in washington state - fieldrobot · robotic apple harvesting in washington...
TRANSCRIPT
Supported By: National Institute of
Food and Agriculture
Robotic Apple Harvesting in
Washington State
Joe Davidsonb & Abhisesh Silwala
IEEE Agricultural Robotics & Automation Webinar
December 15th, 2015
a Center for Precision and Automated Agricultural Systems (CPAAS) & b School of Mechanical and Materials Engineering,
Washington State University (WSU)
1
Supported By: National Institute of
Food and Agriculture
Acknowledgements
This work was funded by the United States Department of
Agriculture – National Institute of Food and Agriculture (USDA-
NIFA) through the National Robotics Initiative (NRI).
2
Supported By: National Institute of
Food and Agriculture
Presentation Overview
• Motivation
• Working environment
• Design objectives
• Hand picking analysis
• System design
• Preliminary field testing results
• Future work
• Questions
3
Supported By: National Institute of
Food and Agriculture
Research Motivation• Washington State fresh market apple industry in 2014
– 2.7 million metric tons of apples valued at $1.84 billion USD1
– Accounted for 70% of U.S. apple production
• The WA fresh market apple harvest requires
– Employment of 30,000 additional workers
– An estimated cost of $1,100 to $2,100 USD per acre per year2,3
• Labor costs are rising and there is increasing uncertainty about the availability of
farm labor
• Lack of mechanical harvesting for fresh market apples is a significant problem
1 USDA National Agricultural Statistics Service. (2014 Washington Agriculture Overview). Retrieved November 2, 2015, from
http://www.nass.usda.gov/Quick_Stats/Ag_Overview/stateOverview.php?state=WASHINGTON
2 Galinato, S., & Gallardo, R. K. (2011). 2010 Estimated Cost of Producing Pears in North Central Washington (FS031E). Retrieved January 13, 2013, from
http://extecon.wsu.edu/pages/Enterprise_Budgets.
3 Gallardo, R. K., Taylor, M., & Hinman, H. (2010). 2009 Cost Estimates of Establishing and Producing Gala Apples in Washington (FS005E). Retrieved January 7, 2013, from
http://extecon.wsu.edu/pages/Enterprise_Budgets.
4
Supported By: National Institute of
Food and Agriculture
Long-term Goal: Reduce dependence on the labor force for
fresh market tree fruit harvesting
5
Supported By: National Institute of
Food and Agriculture
Working Environment
• Commercial apple orchard located in Prosser, WA
• Highly unstructured environment
• Modern cultivation systems with formal tree architectures
• “Fruit Wall” concept simplifies the task6
Supported By: National Institute of
Food and Agriculture
Design Objectives
• Cycle time < 6 sec
• Detachment success > 90%
• Fruit damage < 10%
• Pick multiple apple varieties
• Modular design that is cost-effective
• Our Approach: An ‘undersensed’ design that executes
look-and-move fruit picking, is mechanically robust
to position error, and replicates the human picking
process
7
Supported By: National Institute of
Food and Agriculture
Initial Design Development: Manual Apple Picking
• Fruit is grasped with a spherical power grasp4 with the index
finger applying pressure against the stem
• No dexterous manipulation of the fruit with the fingers
• To separate the apple from the branch, the hand moves the fruit
in a pendulum motion
4 Cutkosky, M. R. (1989) On Grasp Choice, Grasp Models, and the Design of Hands for Manufacturing Tasks. IEEE Transactions on Robotics and
Automation 5 (3): 269-279.
8
Supported By: National Institute of
Food and Agriculture
‘Undersensed’ Hand Picking5
• Are there effective methods to pick fruit that do not require fruit
orientation and stem location?
9
5 Davidson, J., Silwal, A., Karkee, M., Mo, C., & Zhang, Q. (2015). Hand Picking Dynamic Analysis for
Undersensed Robotic Apple Harvesting. Transactions of the ASABE. (Under review)
Mechanical Design
Supported By: National Institute of
Food and Agriculture
• Custom design
• 7 degrees of freedom
• Modular configuration
(Dynamixel Pro
actuators)
11
End-Effector Design
Supported By: National Institute of
Food and Agriculture
• Underactuation provides shape-adaptive
grasping
• Passively compliant joints enhance
robustness to position error & unplanned
collisions6
• Grasping is executed in an open-loop
manner
• Fabricated with additive manufacturing
12
6 Davidson, J., Silwal, A., Karkee, M., & Mo, C. (2015). Proof-of-Concept
of a Robotic Apple Harvester. Robotics and Autonomous Systems. (Under
review)
Vision System7
Supported By: National Institute of
Food and Agriculture
Color CCD Camera
PMD Camcube 3.0 (ToF, 3D camera)
IdentificationFusion
1
2
3 Localization
Camera Rig
13
7 Silwal, A., Gongal, A., & Karkee, M. (2014). Identification of Red Apples in Field Environment with
Over-the-Row Machine Vision System. Agricultural Engineering International: Agric Eng Intl (CIGR
Journal), 16(4), 66-75.
Experimental Setup8
Supported By: National Institute of
Food and Agriculture14
8 Silwal, A., Davidson, J., Karkee, M., Mo, C., Zhang, Q., & Lewis, K. (2015). Design, Integration and Testing of a Robotic
Apple Harvester. Journal of Field Robotics. (To be submitted)
Vision Performance
Supported By: National Institute of
Food and Agriculture
Actual vs. Recovered Image17
Task Timing & Vision Accuracy
Supported By: National Institute of
Food and Agriculture
Vision Accuracy:
Total # of Images = 54
Total Fruit Manual Count: 193
Total Fruit Identified: 193
Identification Accuracy = 100%
Total Fruit in Workspace = 150
Average Fruit per Image = 4
Average Vison Time per Image = 6.3 s
Average Vision time per Apple = 1.7 s
18
Picking Results
Supported By: National Institute of
Food and Agriculture
• 127 of 150 fruits attempted were picked (approximately 85%)– 8/127 – No stems
– 33/127 – Spur attached to fruit
– 86/127 – Stems attached to fruit
• Misses fall into the following five general categories1. Poorly thinned branch (aka “fruit pendulum”) – 7 instances
2. Finger grabbed adjacent obstruction – 3
3. Position and/or calibration error – 8
4. Fruit slipped from grasp – 2
5. Previous fruit stuck in hand - 3
• No obvious evidence of bruising
• Ideal fruit location is 3 – 6 in away from the trellis wire
19
Picking Time
Supported By: National Institute of
Food and Agriculture
• Mean picking time – 6.01 sec/per fruit– 1st fruit in a cycle: 6.22 sec
– Remaining fruits in a cycle: 5.84 sec
• Each task in the picking sequence was segregated into an
individual function and timed
– Motion planning computation: 0.15 sec
– Approach: 2.14 sec
– Grasp: 1.5 sec
– Removal: 1.23 sec
– Fruit release: 1 sec
0
0.5
1
1.5
2
2.5
Motion
Planning
Approach Grasp Removal Fruit Release
Picking time
Tim
e
20
Future Work
Supported By: National Institute of
Food and Agriculture
• Higher level decision making based on detection of
trunks and trellis wires
• Grasp planning based on visual input
• Tactile sensor integration for detection of stem break,
missed fruits, etc.
21