By spring 2015 my company, VC-backed Harvest Automation, was in an awkward situation. Our market was not growing fast enough and our attempted pivot wasn’t getting good traction. But in neither case, did it appear that a clever technological innovation could rescue us. Since technology was pretty much all I did (I was Chief Technology Officer after all) I started thinking that maybe I should look for something else.
It was about that time that one of my favorite conferences took place. Sponsored by Xconomy, and held at the Google offices in Cambridge, MA it was called Robo Madness. I liked Robo Madness because there were always interesting talks about robots and I invariably ran into people I knew. The 2015 conference didn’t disappoint me as I encountered a recent acquaintance of mine, Dave Dorhout. We began chatting about a project he was working on. A budding entrepreneur from Iowa, Dave had started experimenting with mechanisms that might enable him to build a weeding robot for home gardens. I was instantly intrigued.
Farm to Garden
Harvest had perfected a robot that “spaced” potted plants on nursery farms (see Harvest Automation, Part 1.) But a single product offering (or SKU in business parlance) is rarely enough for a company, so we searched for other products in the same domain. Probably the most frequently suggested agricultural robot was one that could eliminate weeds without using herbicides. So why wasn’t Harvest working on that? Because real world constraints make it deceptively and excruciatingly hard to pull off.
In 2015, several robot companies were trying to develop weeding robots. But none had produced a practical machine or secured a profitable niche in the marketplace. Try as I might, I’d been unable to come up with any ideas that seemed better than the ones other companies were already working on.
But as soon as I considered it, I realized that changing the venue simplified the problem tremendously. The reduction is scope was huge—rather than being responsible for weeds in fields that might encompass tens or hundreds of acres, a home gardening robot would typically need to cover less than 1% of one acre. Distinguishing between weeds and crops should also become easier. If needed, a gardener might be willing to attach some sort of identifying tag to each of his or her say, 50 plants. But a farmer would balk at tagging maybe 30,000 plants per acre.
Dave’s idea, of pursuing home gardens rather than farms, might make the robot weeding problem tractable. A garden robot wouldn’t interest Harvest Automation. We needed a big market like commercial agriculture to make our economics and existing sales channels work. But a new, smaller company might find its fortune in home gardens. Dave offered to show me what was working on.
A few days later I trekked down to visit Dave in his Waltham apartment. Although it was early in his development process, the potential of his idea was unmistakable. “I want in!” I told Dave. He graciously agreed to let me join the project. There was much to do.
A marketable product generally requires the harmonious confluence of several disciplines. At a minimum, a robot has mechanical, electrical, and software needs. Neither Dave nor I had strong electrical engineering skills so, I set out to recruit an old friend of mine who did. A capable fellow named Jay Francis and I had worked together at iRobot in the early days long before Roomba. Happily, Jay was available and up for an adventure; he agreed to join us.
The first order of business, as we saw it, was to engineer an effective way to remove weeds from the ground. We considered various types of augers that could dig them out, counter-rotating cylinders that could grab weeds and pull them out, and I even worked on a solarization scheme. The latter idea would have the robot carry around a small, clear plastic dome that it would place over a weed on a sunny day. The sun would cause the temperature of the air trapped inside the dome to rise to the point that it killed the weed. Relying on the sun would let the robot eliminate weeds using very little power.
But on a robot, all systems need to work together. A mechanism that’s great at accomplishing one part of the problem may be disqualified if it relies on another mechanism that’s impractical to build. That turned out to be the case with many of the weed-killing ideas we came up with. To be effective, the solarization, roller, and auger methods all needed accurate information about the position of the weed. But identifying a weed versus a crop plant was not a solved problem at that time.
We continued to experiment until Jay came up with something interesting. He built a little two wheeled remote-controlled device and added to it a tiny cutting mechanism that resembled a weed whacker. (For mysterious reasons, Jay liked to add googly eyes to all his robots.) We all got together at Jay’s house to try it out. It worked wonderfully! The other mechanisms we’d tested earlier required careful positioning and they removed weeds slowly. But Jay’s device was speedy and quickly macerated noxious weeds most satisfyingly. Whacking was the way to go.
Weedebud
Soon after his whacker discovery, Jay flew across the country to attend one of his favorite events: The Bay Area Maker Faire in San Francisco. During the last hours of the event he ran across a curious exhibit with a banner proclaiming: Weedebud — affordable home gardening robot. This robot, the poster said, used AI to keep weeds at bay in gardens and flower beds.
Jay took numerous photos and emailed them to us. Viewing the pictures, my heart skipped a beat.
We believed the time had come for a garden-weeding robot, but we hoped we’d be first off the blocks. It was unnerving to discover that another group was already demonstrating a cool-looking prototype while we were still wrestling with basic concepts. Were we beaten before we even began? We learned everything we could about our garden rival.
Overall Weedebud was small, as we expected our robot would be. Weedebud had two tall drive wheels on either side and a solar panel on top. In front was a camera mounted low to the ground. On the bottom, a small feature like the claw of a claw hammer served to pull weeds. Repeatedly we watched Weedebud’s online video. It showed the robot moving about a garden, positioning the claw to extract a weed and then… the video stopped.
Jay expressed our suspicions, “I don’t think it works.” He said, “Otherwise why wouldn’t they show the full action?”
Anyone who works in robotics learns that building a cool prototype and devising a wow-worthy demo is the easy part. Translating your concept into a product people will buy is the real challenge. That demands a rarer set of skills.
Taking that measure of our rival we wondered: Did Weedebud have a dependable algorithm able to distinguish weeds from garden clutter? Were the AI-enabling camera and processing hardware affordable at a consumer price point? Once a weed was identified, could the robot maneuver autonomously with sufficient reliability to snare the weed with its narrow claw? Did the claw actually work? We couldn’t find evidence that Weedebud had mastered the many practical challenges required to succeed as a consumer product.
Normal cardiac output resumed.
(Historical note: Our rivals changed the—potentially misleading—name of their robot from Weedebud to Weedobot. But despite that appellation advantage, the robot never reached market.)
Reboot
Unfortunately, Dave’s circumstances unexpectedly changed. Our collaboration ended and he soon began working full-time at CSAIL (previously known as the MIT AI Lab). I was sure Dave would enjoy his time there as much as I had when I worked at the Lab many years earlier.
Having convinced ourselves of the garden weeding robot’s potential, Jay and I were determined to carry on. But with Dave gone we needed someone to perform all the CEO duties that Jay and I loathed.
At Harvest Automation, my previous company, I’d worked with a highly talented engineer/MBA, Rory MacKean. Rory had tackled many, varied jobs at Harvest and excelled at them all. In the small world of mobile robotics, Jay had also worked with Rory at a different company (MobileRobots, Inc.). Jay was as impressed as I was.
Rory had only recently left Harvest Automation for a position at a company called Superpedestrian. But heedless of that, Jay and I begin a relentless recruitment drive. Rory’s new company was fully funded—providing him with a reliable salary. His office and lab were well appointed, and the company’s highly anticipated product was on the cusp of launching.
Our counteroffer to those perks were: no salary, no office, and a product with almost nothing resolved—but a really fun project and a chance to build world-changing robot. A few weeks later, over sous vide salmon at his house Rory relented, “OK, you’ve worn me down, I’m on board.”
(Aside: Have you noticed in my stories that robots seem to share some qualities with powerful drugs? No matter how many times they deliver a bad trip, some of us just keep coming back again and again.)
Tertill
Names are important but tough to do well. To seek advice on naming our new robot I arranged to meet with my colleague from iRobot, Nancy Dussault Smith, the most talented marketer I knew. Nancy was responsible for much of the early publicity that made Roomba a household name. I told her what we were planning; she knew what we should do. Nancy said, “Give the robot a cute name, maybe something like Carrot.” I was sure Nancy was on the right track but calling our robot Carrot would seem to play favorites among the vegetables. So I looked for a moral equivalent.
Turtles are small, cute, and like to spend time in gardens. Also, there have been several influential turtles in robot history; for example, Seymour Papert’s Logo Turtle and W. Grey Walter’s “terrapins,” Elmer and Elsey. But to stand out we needed a different spelling. By combining the garden-related words ‘terra’ and ‘tillage’ Tertill got its name.
Now all we had to do was pretty much everything else—solve a plethora of technical problems, find funding, implement manufacturing, and master marketing, sales, and distribution.
