Harvest Automation, Part 1

[Part 1, Part 2, Part 3, Part 4] 

After leaving iRobot in November 2006, three colleagues and I founded Harvest Automation, Inc.  Over several years we researched a new application, secured funding, and built a robotic product that worked as we intended.

Founding

In early 2006, having spent 10 years as the General Manager at DEKA, Charlie Grinnell joined iRobot.  At the time, I and a few others had been working to initiate an advanced development group at iRobot; Charlie was appointed our leader.  

Trained as a civil engineer, Charlie spent the bulk of his career on the management/business side of enterprises.  In contrast to the experiences I’d had with other practitioners of those disciplines, I found my association with Charlie to be refreshingly positive.   Charlie listened to creative people, respected the creative process, and did not play power games.  We got along splendidly.

But “splendid” did not describe the relationship between Charlie and his immediate boss.  After about six months our new leader was let go and our advanced development group was set adrift.  

It was at this time that several of us reluctantly concluded that iRobot was no longer interested in revolutionary products.  And that, if we wanted to find the next big thing in robotics, we’d have to start our own company.  Besides me, our impatient band included mechanical engineer Paul Sandin and software engineer Clara Vu.  Clara and Paul had worked with Charlie too, and we all agreed that our first step should be to woo Mr. Grinnell to become our CEO.  After numerous lunches, meetings, and visits to Charlie’s grand old house in Groton, MA we succeeded!  

To ease our departure from iRobot, we sought the assistance of Frank Ingari, a grizzled computer industry veteran, who was then advising iRobot.  With Frank greasing the skids, we were able to leave en masse on the friendliest of terms.  As iRobot Chief Council Glen Weinstein told us, “We want you to succeed, just not in consumer floor cleaning.”  We were fine with that, certain that the next big thing in robotics lay elsewhere.

The first order of business was to name our newly minted startup.  After a much too drawn out search we settled on Q Robotics.  Why Q?  The nerdy reason is because Q is the figure of merit of a tuned circuit.  We were interested in building, not general-purpose robots—which we all felt made an unkeepable promise—but rather robots tuned to a particular task.

At Q’s founding we had no external funding, no customers, and no particular product in mind.   Our assets consisted of some savings, strong technical chops, and abundant enthusiasm.  After all, we’d just bested scads of competitors by delivering the world’s first wildly successful consumer robot.  How hard could it be to do something like that again?

Product

What revolutionary new robot should we build?  We left iRobot having deliberately given that question very little thought.  This approach was  necessary, we believed, because if we developed a concept while employed at iRobot, the company might claim ownership of our idea.  

We set off in search of the right concept.  The trick, we agreed, would be to find an application where the needs of the task matched well with the abilities of a robot.   Roomba had succeeded at consumer floor cleaning—the lowest hanging fruit on the tree of robot possibilities—what might the next more difficult problem be?  Clara liked to say, “The application is the solution.”

We searched extensively while purposely excluding several popular options.  We saw “elder care” robots, for example, as far too ambitious.  Lawn mowing robots were already being explored by the Swedish company, Husqvarna, and robots intended to entertain or socialize left us unmoved.  The latter’s appeal, we thought, would last only until the novelty wore off.  We were all strongly of the mind that robots should preform useful, physical work—not sing and dance.

We brainstormed, we scoured the web, and we visited sites where robots might prove useful.  Paul and I, for instance, paid a call to a small food distribution warehouse looking for ideas.  We found nothing suitable there but, from all our efforts a few fanciful robot possibilities emerged:

• MooDroid – A robot that collects manure on a farm and delivers it to an energy-producing methane digester.
• DooDroid – The cousin of MooDroid, this robot removes doggie doo from the lawn.  (Later, another company tried to build such a robot.)
• DruidDroid – A robot that maintains the ground cover at Christmas tree farms.
• NoroBot – This robot constantly swabs doorknobs and takes samples on cruise ships looking for contagions. 
• Aqua Cine – A training robot for competitive swimmers.  The robot drives along the bottom of the pool tracking and filming the swimmer on the surface for later analysis.
• BoatBottomBot – This robot performs the disagreeable task of cleaning the hulls of pleasure boats.

We spent much time analyzing and debating these and other potential products.  All seemed feasible but none felt compelling.  Then in February 2007, still searching for our grail, we learned of a trade show called New England Grows to be held in Boston.  Maybe some niche in agriculture held promise for robots, we thought.  Paul and I arranged to attend.

At the show, we happened to speak to a representative of Cavicchio Greenhouses Inc. who told us an intriguing story.  Every nursery and greenhouse (N&G) grower in the world, he said, had the same problem and no one knew how to solve it. 

The problem related to producing plants in pots.  When the plants are small, the pots are packed tightly together on growing beds.  Such packing minimizes the grower’s need for (expensive) land and makes watering efficient.  But as the plants grow, a point is reached where the leaves and branches from different plants begin to interfere with each other.  At this stage, in order to avoid damage, the plants must be moved apart so each has room to grow. 

The operation that accomplishes this is called “spacing.”  Given that one to two billion potted plants are grown annually in the US, a literal army of plant-moving people is needed.   Cavicchio invited us to visit their farm where we witnessed a crew of eight spacing plants.  The workers’ job was physically demanding, and injuries were common. 

The idea of a plant-moving robot instantly attracted us.  We thought that if we couldn’t build a robot that could pick up a pot, carry it a few feet, and put it down again, we were in the wrong business.  That robot, initially called PotBot, leapt to the top of our potential products list.  Contributing to our interest was the fact that N&G farms constituted a $17 billion industry and by far, labor accounted for the largest fraction of their costs.  That gave us a path for current and future advancement.  We could start with moving plants and then build more labor-saving robots as we better learned the industry.  

Proof-of-Concept

It took PotBot less than a month to sweep away all competitors.  A product that would perform the spacing operation at nursery farms and greenhouses would be the next big thing in robotics and we would make it happen.

To stretch our modest funds, rather than renting an office, we set up our company in Charlie’s billiard room.  We exiled the pool table, moved desks in, attached white boards to walls, and hauled in tools and computers.  Cavicchio had given us a bunch of sample pots to experiment with, and we began thinking about how to make our robot pick them up and put them down.  Charlie started contacting other growers to arrange visits, talking with the N&G trade organization, and planning the business side of our assault on the industry.

The new robot would necessarily be more complex than Roomba.  Navigation by just bouncing around wasn’t possible in this application.  At least some of the time the robot would need to know where it was so it could put plants down in a space-efficient pattern.  It would need an arm and gripper with which to manipulate pots.  And the robot would require sensors to enable it to find a pot to pick up and to identify the pattern where it would place the pot that it carried.  

Building such a robot would require a bigger development effort than Roomba and we knew our meager savings would be nowhere near enough.  We’d have to get significant funding from deep pocketed benefactors.  How would we convince investors that our idea wasn’t just a roboticist’s fever dream?  Obviously, we’d first need to build a proof-of-concept robot to show them the soundness of our approach.  (Later, we’ll find a good bit of irony in our “obvious” plan.)

So the four of us (Charlie, Clara, Paul, and me) spent around a year building a proof-of-concept prototype of the robot we hoped investors would fund.  We named it Zeke.  It had two large drive wheels, and a low-friction slider in the front for balance.  It used a pair of automotive window motors to move its arm up and down.  Mounted low across the front was an array of infrared range sensor that let the robot identify and home in on pots.  Standard DeWalt rechargeable batteries supplied Zeke’s power.  

Zeke relied on dead reckoning to keep track of its position.  That let us initially skip some of the development the final product would require, but it meant the prototype couldn’t do the whole job.  Still, we thought it would do enough to enable a promising demo.  Ultimately, we built a robot that could select one pot from a group of pots, pick it up, drive a few feet with it, put it down, and then go back for another pot.  

Pitch

The presentation instrument (often a PowerPoint slide show) that funding supplicants prepare for investors is called a “deck.”  We constructed our deck (tweaking it continuously throughout our funding marathon), packed up Zeke, and took our show on the road.  Our timing could hardly have been worse—we began looking for funding just as the financial crisis of 2008 (remember credit default swaps?) began to devastate the investment market.  Finding investors devolved into an interminable, frustrating slog.  

We pitched our plan to probably 50 different venture capitalists, VCs.  (That’s a lot.)  It didn’t help that our business category was “robots for agriculture.”  That combination seemed to push most VCs well out of their comfort zones.  Very few we talked to were familiar with robots, and almost none had any understanding of agriculture.  But we carried on.  When we were down to the fumes of our savings things finally fell into place.  MassVentures (then called MTDC), a VC fund established to help early-stage startups in Massachusetts, was first to commit their cash.  Joining them were Cultivian Sandbox Ventures (a company specializing in agriculture),  the investment arm of Syngenta (a global agricultural company), and  Founders Collective (a group bold enough to fund something unfamiliar).  

Our quest for cash had taken more than a year longer than we had anticipated but at the end of 2009 we closed on a deal that put several million dollars into our company’s bank account.  We moved into a real office and started hiring a team to develop our robot.

Once it was clear which industry we would serve and what our product would be, we changed the company name to better reflect our focus.  We abandoned the generic Q Robotics in favor of the more specific Harvest Automation.  And we gave our robot a name worthy of a piece of farm equipment,  HV-100 (sometimes referred to as Harvey).  

The ironic part of the story:  We thought a prototype was essential because we expected lots of pushback from potential investors.  Surely, they’d worry that we couldn’t deliver something as audacious as a pot-moving robot.  (We certainly would have doubted any technology team who made that claim!)  But it turned out that VCs had no such concerns.  They regarded us as the technology experts, if we said we could build a pot-moving robot, they believed we could.  The aspects of our pitch that investors objected to were always things like the size of our market and how quickly our product might be adopted.

 So, if instead of spending a year (and our savings) building a sophisticated prototype, we had simply approached investors with our big idea, we might have gotten funded in the flush times before the crash.  And that would have let us begin work on the robot nearly two years sooner.

Development

There was much to do:  We’d have to find a way to position the robot precisely.  The robot would need a very robust design able to handle mud, dust, and bumpy surfaces.  A mechanical lifting mechanism that could operate reliably for thousands of cycle was necessary.  And we’d have to devise a simple/cheap way to let customers input the task’s  parameters.  (The robot had to be told things like the desired pot-to-pot distance and the pattern; e.g., hexagonal vs rectangular.)

One of the most interesting innovations we devised, I thought, was the way we kept the robot registered to the growing bed.  In the time before AI exhibited skill at identifying images, we couldn’t use a camera.  Rather, we needed a reliable, non-visual way for the robot to recognize the edge of the growing bed and for workers to tell the robot where they wanted the pots placed.  

We came up with a strip of retro reflective material (like road signs and safety vests use) that the operator would lay down along the edge of the bed.  Then we built sensors that could very accurately follow that strip. The robot would start near where some unspaced pots had been dropped off onto the bed.  It would pick up a pot and turn toward the strip and drive forward until it detected the strip.  Next it would follow the strip until it saw a collection of spaced pots in front of it.  The robot “saw” the pots using an advanced scanning laser rangefinder made by the German company SICK.  

I hated the SICK.  It was a relatively delicate instrument and terrifyingly expensive at around $3000!  After the frugality of Roomba, paying that much for one sensor was a body blow.  We searched mightily for an alternative.  We researched other manufacturers, we tried to build a sensor of our own, we investigated stereo cameras.  But ultimately, we had to adopt the SICK.  It could see the black pots growers commonly used even in bright sunshine.  At the time, no less-expensive technology could reliably do the same. 

After close to four years of development we finally had a product to sell.  Our robots generated a great deal of interest within the N&G industry.  Harveys appeared in trade magazines, and on a few news shows.  They always drew a crowd at N&G conferences.  We talked to many growers and conducted demonstrations at numerous farms and greenhouses around the country.  At our peak we had maybe 150 robots in customers’ hands that together moved several million potted plants per year.

We charged about $30,000 per robot and we resisted selling them in bundles of fewer than four machines. Customers also bought service contracts.  A good bit of cash flowed into the company.

Our prospects were positive.