Robots will grow our food.  Or so futurists have predicted since the Art Deco era.  And the vision is beguiling.  In principle, robot farmers could control weeds without herbicides, they could maximize yields by attending to the needs of each plant individually, and they could harvest produce without human toil or injury.  Laudable goals all.

But robot farming is tricky.   Although many companies have presented impressive demonstrations of robots performing various farming tasks, none of these new techniques have displaced conventional methods in the marketplace.  For instance, despite much excitement around robotic crop weeding, global herbicide production continues to increase.  

During my nine years at Harvest Automation I frequently searched for ways to expand our product offering.  Our HV-100 robot catered to the nursery and greenhouse industry and that had given us insights, credibility, and sales channels in agriculture.   Clearly, the sort of new product that would fit most naturally into our business would be another ag-focused robot.  Also, as a result of our work, I’d developed a strong appreciation of agriculture.

Agriculture appeared to offer many possibilities for robots.  As mentioned, eliminating weeds without herbicides was one attractive example.  So, I considered building a weeding robot for farms, but I was late to that party.  Many other companies were actively working on the challenge of robotic weed control but, at that point, without compelling success.  Although it’s relatively easy to create a demonstration of a robot removing weeds from a field, it’s another matter entirely to deliver a robust and effective weeding product at a price farmers are willing to pay.  I was not able to come up with an idea for a weeding robot that I considered to be head and shoulders above what others were already attempting.  

So I looked elsewhere.

Many major crops like corn, wheat, and soybeans are harvested mechanically.  When the crop is mature, a combine harvester or other implement drives through the field.  The machine efficiently harvests all the plants at once leaving behind only stubble.  But other crops are not so obliging.  Delicate crops and those that ripen in stages, raspberries for example, cannot be treated so rudely.  Workers must return repeatedly throughout the season to harvest the crop.   And they must gather the fruit without damaging the plant.

It turns out that hiring people to pick crops is vastly more expensive and logistically challenging than driving a mechanical harvester through a field.  An advisor Harvest Automation worked with related a telling anecdote to us.  Our advisor managed strawberry picking operations in California, while his brother in the Midwest was responsible for a 9,000-acre corn farm.  Extensive mechanization enabled the brother to run his entire operation with only ten workers.  That is, each worker could maintain 900 acres.  But our advisor needed to hire one to two workers for each acre of strawberries (mostly for picking).  On average then, it took over 1000 times as many people to maintain and harvest strawberries as corn.

Hand-harvested crops thus seemed ripe (so to speak) for robotization.  But building a single robot able to harvest any sort of crop didn’t appear viable to me.  General solutions are almost invariably more expensive and difficult to engineer than particular solutions.  I expected that an economical robot would have to specialize.  But which crop should I choose?

If a different type of robot is needed to harvest each sort of crop, then one attractive approach is to select the crop with the greatest market value.  That ought to deliver the largest economic reward to the robot builder.  At over $2 billion in annual revenue, a strong case can be made for choosing strawberries.  Of course, the reward is available only if the robot actually works.  And sadly, to date, no attempt to build a strawberry picking robot has made a major dent in the market.

Filter

I decided to approach the problem from the other direction and consider which crop would give the robot the best chance of technical success.  That is, of all hand harvested crops in the US, which ones were most amenable to robotic harvesting?   

One lesson I’ve learned is that in order to build a good robot you should always start by trying not to build a robot.  If a task can be accomplished by a mechanism that’s simpler and lower in cost than a robot, then build that mechanism. Simpler, cheaper will beat the robot every time.  So, in my survey of crops, it was important to avoid choosing a crop that might soon be harvestable by non-robotic means.

I constructed a system to score crops according to their potential for robotic harvesting.  Of the many features crops possess, I settled on these six as the most significant: plant preservation, obscuring foliage, ripening uniformity, ease of identification, fruit altitude, and toughness.  

Plant preservation

Must the plant survive the harvesting of its fruit?  Crops like corn and soybeans are replanted each year, so it’s permissible to simply clear-cut the growing fields—destroying the plant as you harvest the produce.  But other crops, avocado for example, must be treated gingerly.  The fruit of such plants must be picked carefully, leaving the plant undamaged.  Robotic harvesting is favored if the plant must be preserved.

Obscuring foliage

Foliage can conceal the fruit of some plants, for example strawberries and blueberries.  The more foliage tends to hide the fruit, the less favorable that crop is for robots.  To locate hidden fruit, the robot may have to push or blow foliage out of the way.  That adds complexity to the design and may reduce yield because some fruit will evade the robot.

Ripening uniformity

If all the fruit on a crop plant ripens at once, then a simple one-pass mechanical solution may suffice.  But if the fruit matures in stages, then the harvesting mechanism must make a nuanced assessment, pick some fruits, and leave others to continue maturing.  Non-uniform ripening favors robots over mechanical means. 

Ease of identification

Is it easy or hard to identify the ripe crop?  If the fruit is the same color or shape as the foliage or other part of the plant, or if ripeness is determined by subtle changes in hue, then the robot will find it more difficult to accurately select a target for harvesting.  This may result in false positives or false negatives.  Where, respectively, a part of the plant that shouldn’t be picked is, or a part that should be isn’t.  Easy identification simplifies robotic harvesting.

Altitude

How far above the ground is the fruit positioned?  If it’s high in a tree like apples, then the robot will likely be large and must employ a long arm, possibly with multiple degrees of freedom.  Large, complex appendages make a harvesting robot expensive.  Low-to-the ground fruit means that the robot can likely be smaller, simpler, and less costly.

Toughness

How easily damaged is the fruit?  Is it delicate like strawberries or tough like artichokes?   The tougher the fruit the easier it will be for the robot to harvest it without damage. 

Results

I  examined all the hand harvested crop for which I had information (there were 42 on my list).  For each plant, I assigned scores to each of the six categories and summed the numbers.  My analysis would necessarily be rough, so I allowed a category to have only one of three values.  Zero meant unfavorable to the robot, two was favorable, and one was in between.  Of all the crops I examined, it surprised me that asparagus (green, not white) came out on top with a score of 11 out of a possible 12.

Asparagus is good for robots for these reasons:  Asparagus is a perennial.  After planting, it may take three seasons to produce usable spears.  Then, it continues producing for 15 or more years.  That makes it very important to preserve plant (specifically the underground “root crown”) from year to year.

The part of the asparagus plant that’s harvested is the stalks or spears.  Spears grow prominently, not surrounded by any sort of foliage.  So, the robot’s view of its target is unimpeded.   

Spears continue to emerge and grow over the entire season.  For manual operations, this requires that workers harvest the new growth every few days.  A chore robots would gladly take over.  

Determining the ripeness of a spear could not be easier: height equals maturity.  Tall spears should be harvested, short ones left to continue growing.  

Asparagus spears don’t grow on trees but rather directly from the ground.  Thus an asparagus harvesting robot need not be massive or have a long reach.

The only point asparagus loses is in the toughness category.  I gave it a one rather than a two.  It’s tougher than strawberries but less robust than say, pineapple.

On the whole then, robots like asparagus.  My analysis concluded that it should be easier to build a robot to harvest asparagus than any other currently hand-harvested crop.

Economics

Of course technical matters are only one component of a successful robotic product.  Others include need, competition, and the availability of funding for development.

In current asparagus harvesting practice workers walk the field, stoop over, and then use a Y-shaped knife to cut the asparagus stalks one at a time just below ground level.   The spears may then be placed in a container strapped to the worker’s waist.  The job is harsh, and growers sometimes have trouble finding willing laborers.  There are stories of asparagus being allowed to go to seed or being plowed under because too few harvesters could be found. 

Driven by its great taste and health benefits, consumption of asparagus in the US has generally been expanding.  But while US consumers eat more and more, US growers produce less and less.  One reason for this mismatch is the high cost and uncertain availability of workers to harvest asparagus.  Robots could reduce the cost of harvesting and ensure harvesting happens on time.  That might enable some production of asparagus to return to the US.

Funding is where the story falls apart.  Domestic production of asparagus averaged about $70 million in the past few years.  (We consumed at least five times this amount, the balance supplied by imports.)  Developing and productizing a robot to harvest asparagus is time consuming and risky.  At $70 million, the size of the asparagus opportunity is just too small for most investors.  It’s hard to generate interest among venture capitalists (VCs) for markets of less than $1 billion.

Muddy Future

I wrote an earlier version of this story for my Practical Roboticist blog in 2015.  That story ended at this sad point.  But there have been subsequent developments.  A few years ago I was thrilled to discover that a company called Muddy Machines had been established in the UK with the goal of building a robot to harvest asparagus!  (The founders even consulted me for some advice at the time.)

Muddy Machines did a good bit of development work quite quickly.  They attracted the interest of a major asparagus grower, built an impressive robot, and demonstrated it harvesting asparagus in farm fields.  But this story ends even more sadly than my earlier one.   Muddy Machines was unable to secure the funding they needed to develop their robot into a viable product and are currently winding down.

The nature and size of the market create the funding conundrum.  Asparagus growers might reasonably bankroll the development of a harvesting robot themselves, but their thin margins and risk aversion work against this.  The traditional source of high-risk investment, VCs, are turned off by the small size of the market (although, a successful harvester for one crop could lead to many others).   And an appeal to crowd funding is complicated because the product itself doesn’t directly benefit supporters.  It may be that finding a motivated angel investor offers the best chance for funding an asparagus-harvesting robot.

Disregarding the daunting impediments, another company called Autopickr, also in the UK, has assumed the asparagus-picking mantle.  Good luck to them.  They will likely need it as no aspect of agricultural robotics is easy.  But hopefully, a successful robotic asparagus harvester will emerge in less than the four decades it took to get from the first serious attempt at a robot vacuum cleaner to Roomba. Fingers crossed.