—Ross Courtney
A 3D-printed end effector for a robotic apple picker can be seen plucking green apples from branches during a demonstration by Washington State University bioengineering students near Prosser, Washington, in June. The university’s Center for Precision Automated Agricultural Systems (CPAAS) is aiming to design a robot that can handle a variety of apple-related tasks by switching between tools on the end of its robotic arm. (Ross Courtney/Good Fruit Grower)
Every year, the apple thinning season overlaps with the cherry harvest, forcing us to make difficult labor management decisions.
Northwest engineering researchers are working on developing a bot to do just that.
Students and researchers from Oregon State University and Washington State University are developing a robotic fruit picker that uses a computer vision-guided robotic arm to identify fruitlets and extend a claw-like end effector to pluck them from the tree and drop them to the ground.
In fact, the engineering team is working on developing robots that will handle a variety of apple functions, including pruning, pollinating each cluster, picking flowers, and picking fruitlets, all using similar arms with six-point articulated motion guided by computer vision and machine learning.
These initially started as separate projects with different funding sources but have been combined in recent years into a multi-purpose orchard robotics system, said Manoj Karki, director of Washington State University at Prosser’s Center for Precision and Automated Agricultural Systems, or CPAAS for short.
These management tools can also be incorporated into harvesting robots by swapping out the end effectors, and a multi-purpose robot approach could ultimately make it easier for growers to justify the cost of expensive machinery, Kirkky said.
Pollinating robots have end effectors that identify flower clusters and distribute pollen, flower-picking robots do the opposite by crushing flower clusters with tiny brushes, and pruning robots sense tree branches and… prune them.
While the researchers are still in the early stages, they hope to eventually partner with a private company that can incorporate the technology into commercial machines. Kalki and his team have experimented with equipping one of the harvesting machines in commercial development with an end effector.
Kirkey estimates his lab has spent about $3 million over the past decade in grants from the Washington State Fruit Tree Research Commission and the U.S. Department of Agriculture’s National Institute of Food and Agriculture.
Thinning out the fruit
The robotic fruit picker has an articulated arm that rotates and bends to reach identified apples, and a two-fingered 3D-printed plastic claw to pluck them off the tree. A video camera sensor with machine learning capabilities identifies apples with a high probability and steers the arm.
It will be many years before the machine is commercially available.
At a demonstration at Good Fruit Grower in early June, graduate students meticulously lined up the machine and debugged the software before the harvest, with the claws sometimes missing the right apples and once removing a thorn.
From right, Washington State University doctoral students Ranjan Sapkota and Martin Churbiya, and visiting researcher Zhichao Meng, position the robot for a thinning demonstration. (Ross Courtney/Good Fruit Grower) The robot arm’s camera uses probability to identify the apple, which can be seen tagged with the software’s default “person” tag. A probability of 1.0 means the computer is 100% sure it has found an apple. The researchers then reprogrammed the machine to label the apple as an “apple.” (Ross Courtney/Good Fruit Grower)
On a scale of 1 to 10, Karkey puts the technology’s readiness level at about 3. Robotic pollination, flower picking and pruning are a bit further along, between a 3 and 4.
But perfection isn’t the real goal. The final machine has to be good enough to provide a reasonable return on investment. A grower might be happy if the robot could handle, say, 90 percent of the work well. As more features are added to the machine, that threshold goes down.
OSU Involvement
At OSU, researchers are using their expertise in developing digital and physical “proxy” environments to train robots throughout the year.
Relying solely on real-world experiments, data collection and training would take only a few weeks to a few months per year, said Joe Davidson, assistant professor of robotics and director of the Institute for Intelligent Machines and Materials in Corvallis.
Davidson and robotics professor Cindy Grimm built a fake tree made of springs and magnets and taught the robot’s end effector to mimic the movements of a human hand picking a real apple. The team also created a virtual orchard and trained computer algorithms for a “perception pipeline” that lets the robot recognize and understand tree structure from sensor images.
Davidson said the company plans to field test these algorithms later this year.
Over the past six years, Davidson said, the research committee and USDA-NIFA have spent about $1.5 million in additional funding on related activities.
Watch as a robotic arm practices picking apple fruit as part of ongoing research at Washington State University.
Source link
