Startup

Starting a company is fun.  In the early days possibilities are limitless, enthusiasm is at apogee, and the confounding problem soon to torment the team are largely hidden from view.  I love being involved in a startup!  (As long as someone else handles the corporate nuts and bolts.)

Early orders of business for the company Rory MacKean, Jay Francis, and I were starting included finding names for the company and the robot and securing a place to work.  We did not do these things in the optimal order.  We first picked the company name, settling on Franklin Robotics to honor inventor/scientist/polymath Benjamin Franklin.  Next, we chose Tertill as the name of our robot.  Had we been savvy in the ways of marketing, we’d have realized that a small startup should choose the same name for both product and company, otherwise customers may be confused and marketing efforts diluted.

We did a little better with our choice of venue, an office at the Innovation Hub (iHub) a newly established coworking space run by the University of Massachusetts at Lowell. Not long after moving in we attracted the attention of John Chase, a young mechanical engineer fresh out of an MBA program at Northeastern University.  John had grown up on a farm and was eager to apply advanced technology to agricultural problems.  We believed John’s background and training would be most helpful for the product we were designing.

With the team assembled, serious Tertill development commenced.  

Tertill Strategy

Our vision was clear: Tertill would live in the user’s garden, keep it perpetually weed-free, and require zero effort on the part of the gardener.   That bold notion would take a lot of innovation to implement.  

As described in Hatching Tertill, we would rely on a miniature weed whacker to attack sprouting weeds.  But so much more was needed.  How would we power Tertill?  How would the robot eliminate weeds while protecting crop plants?  What would keep Tertill in the garden?  How would it overcome baking sun, soaking rain, mud, dust, rocks, and ruts?  And how could we achieve a price that gardeners could afford?  Nothing would be easy.

Here’s a bit of prior history:  Before Rory agreed to be our CEO, I’d ask Winston Tao, the project manager for iRobot’s Roomba, if he would fill that role.  Ever diligent, Winston sought advice from two technologists whose opinions he trusted: Eliot Mack (Roomba’s mechanical engineer) and Phil Mass (Roomba’s software engineer).  Eliot and Phil considered the Tertill proposition carefully.  They concluded that Tertill would be harder to pull off than Roomba principally because the garden robot must have a price close to Roomba’s while also being waterproof and operating on terrain much more challenging than living room floors.  They informed Winston that Tertill couldn’t be built for the price we hoped to charge.  Winston then prudently declined to take the helm of Franklin Robotics.   (But for a couple of innovations Team Tertill came up with, Eliot and Phil would probably have been right.)

Power

What were our options for powering Tertill?  We might use a charging station like Roomba, or we might employ the sort of wireless system many smart phones utilize.  But a simple and satisfying approach would be a solar panel mounted on top of the robot.

For other robots I’ve worked on, Harvest’s HV-100 in particular, folks often asked why we didn’t use solar power.   The reason was always the same, the robot’s power requirements were such that the sun was insufficient—a solar powered HV-100 would use up all its stored energy quickly, leaving it idly charging for much of the day when it should be working.  

But—after careful calculation—I realized with great joy that Tertill was different.  It could accomplish its task using only the energy available from a top-mounted solar panel.  The problem was easier for Tertill because, 1) the area the robot needed to cover was relatively small—an average garden was only about 100 square feet (9 square meters) and 2) weeds didn’t grow that fast—if the robot failed to find and strike down a growing weed one day, it would be back the next.  

Discrimination

We devised a dirt-simple strategy for differentiating between weeds and crop plants.  In our robot’s view, any plant standing tall enough that it touched the side of the robot’s shell was a crop plant.  Tertill turned aside when it encountered this situation.  But a plant short enough to slide beneath the robot was considered a weed.  In that case the touch sensor on the bottom of the robot would activate the whacker and the weed would be chopped down as the robot rolled over it.

With that straightforward choice we eliminated the need for a camera, the computational hardware a camera would require, and the extensive training an AI-based system would depend on.  Tertill would discriminate based on only mechanics and touch.  

But this choice imposed certain consequences.  What if garden crops were planted from seeds?  The robot would mis-identify these desired sprouts as weeds and attack without mercy.   For this case we developed “plant guards.”  The guards were wire barriers, like tiny tomato cages.  Gardeners would place guards around sprouts and delicate crop plants.  When the robot touched the guard, it knew to turn away. 

Obstacle Detection

Roomba used a mechanical, spring-mounted bumper to detect obstacles it contacted.  But Tertill needed to avoid such complexity.  A bumper that moved relative to the robot’s chassis would cause no end of headaches for an outdoor robot.  How would we shield it against dirt and entanglement with vegetation?  How would we prevent water from getting in at the connection points?  Those seemed like daunting challenges likely to require expensive hardware.  

Fortunately, we found a wonderfully simple, reliable, and cheap solution: capacitive sensors.  Buttons on many appliances and some touch screens rely on capacitance to activate.  All such a sensor needs is to come close to a mildly conductive object like a human finger or a moisture-filled plant, and detection is assured.  And even better, our robot’s microprocessor plus a couple of dirt-cheap components could implement such a sensor.  It took a bit of fussing to get the details right but once we did, we had an amazingly cheap and reliable sensor.

Our capacitive sensor wouldn’t have worked for Roomba because of the “mildly conductive” requirement. Plaster walls and chair legs don’t qualify.  There was a similar problem outdoors.  Certain things like say plastic garden gnomes also don’t register.  How should we handle such items?

A frequently suggested solution is to just watch the robot’s wheels.  As long as the robot is making forward progress its wheels turn, but when it runs into an immovable obstacle they may stop.  Unfortunately, they don’talways stop.  If the ground is muddy or very dusty, the wheels may slip and keep spinning while the robot pushes futilely against the obstacle.  

But physics suggested another approach: inertia.  We included in our robot a multi-axis accelerometer.   One of its purposes was to alert the robot when it was attempting to climb, descend, or traverse a steep slope likely to make it tip over.  But the instrument could do more than that.  When the robot sped up, the accelerometer would record an acceleration in the direction of motion, when it slowed down the opposite signal would appear.  So I created a virtual sensor.  Software constantly modulated the robot’s speed—increasing and decreasing the speed about three times per second.   While this happened, my code noted whether the accelerometer responded accordingly.  If it did then all was well, the robot’s motion matched the commands to its wheels.  But if the signal stopped following the modulated wheel velocity it meant the robot was stuck and its wheels were spinning.

The combination of the two methods, capacitive and inertial sensing, gave us a reliable and cheap method do figure out when the robot touched something or got stuck.  

Confinement

Most vegetable gardens are surrounded by fences or some sort of barrier.  This is to demarcate the garden’s boundary and to deter pets from trampling the plants and pests from purloining the produce.  We decided to rely on this common feature to prevent Tertill from wandering away.  The capacitive and inertial obstacle sensors we were building into Tertill would easily detect the boundary.  Any higher-tech solution would have been much more expensive.  We would advise customers on what sort of low-cost fence to use if their garden had none.

Size

The dimensions we chose for Tertill were not arbitrary.  Early in Tertill’s development we surveyed the sort of plants gardeners most commonly grow in their gardens and the recommended plant-to-plant separation. Some crops like carrots and onions may be only three inches apart within the row.  Others like tomatoes and squash are often separated by more than a foot.  But there seemed to be a sweet spot at around eight inches (200 mm) with lots of plants needing this much or more space and only a few needing less.  We decided a practical weeding robot should be no wider than eight inches. 

Shape

We tried to build a rectangular robot.  The form factor of a tractor, with big wheels in back and small ones in front, appealed to us and our second-generation prototype was built this way.   Our hope was that the geometry of plants in gardens would make the environment forgiving enough that we could get away with a non-round shape.  This would simplify the engineering, allowing us to easily pack components into a rectangular box.  

I tried mightily to find a shape and clever algorithm that would give the robot sufficient maneuverability in a real garden.  But it was not to be.  Even simple ubiquitous arrangements of objects—like the corner formed by gardens bounding structures—could easily defeat my best attempts.  We reluctantly concluded that round was the only shape able to give us reliable mobility.  

Propulsion

Outdoor terrain is tough.  Gardens are not flat, and they can contain mud, rocks, ruts, and organic matter.  Tracks or big drive wheels would greatly help traverse such impediments, but our robot had to be small.   We’d been forced to choose a round shape for Tertill.  Other round robots I had built used two drive wheels and roller for balance.  But it turned out that wouldn’t do for our garden robot—such a robot couldn’t reliably climb the modest slopes often found in gardens, say for example, the hills  squash is planted on.  Some physics convinced me that we needed four-wheel drive.

After much brainstorming and testing we ended up with four small wheels cambered (tilted outward) at 45 degrees.  This unusual arrangement let us get good traction and helped the robot avoid high centering.   

Maker Faire New York

In September 2015 a large Maker Faire was held near the site of the 1964 New York World’s Fair.  That gave us our first opportunity to get our concept for a garden weeding robot in front of large numbers of people.  We packed up Rory’s SUV with robots, a garden fence analog, lettuce (to simulate weeds), tools, and computers and headed out. 

Our booth at the Faire was well attended all day long—we had trouble getting away to see anything else.  We demonstrated the robot (the white rectangular prototype) over and over again.  We collected names and email addresses for our mailing list, we talked to lots of gardeners, I shook many hands, and one woman even gave me a hug for having invented Roomba!   Many folks told us that the wanted and needed a robot like Tertill.  It was a great day.  

I was most excited when, at some point in the afternoon, Stephen Wolfram (whom I’d never met before) stopped by our booth.  He took a keen interest in the robot, and we chatted for maybe 15 minutes.  I understood his basic message to be, “You’re doing it all wrong.”  (Stated much more diplomatically, of course.)  Mr. Wolfram asserted that AI image recognition was ready for prime time.  And that that would be a better way of discriminating between weeds and crops than the simple mechanical/tactile system of which I was so proud.  Regardless of our disagreement, it was a fun and enlightening exchange.

***

Much that we did the first two years was building toward an essential market validation step we planned: we wanted to conduct a Kickstarter campaign.  We hoped (and expected) that the response to our Kickstarter would do three things.  1) reveal a widespread interest in our product, 2) provide us with some of the funds we needed for development, and 3) give us ammunition for getting any needed remaining funds from venture capitalists.  Franklin Robotics’ Kickstarter will be the topic of my next post in the Tertill saga.