Roomba Inventor Joe Jones on His New Weed-Killing Robot, and What’s So Hard About Consumer Robotics

The inventor of the Roomba tells us about his new solar-powered, weed-destroying robot

10 min read
Franklin Robotics' Tertill weed-killing robot
Photo: Franklin Robotics

iRobot’s Roomba robotic vacuum is, arguably, the most successful robot ever made. Some 15 million of them are cleaning floors all over the planet, and they’re doing so reliably and affordably and autonomously enough that people keep on buying them, which is something no other consumer robot has ever been able to replicate.

Providing the vision for the small team that designed the Roomba was Joe Jones. What started out as his personal side project at MIT’s Artificial Intelligence Lab in 1988 became a commercial product at iRobot in 2002, and while iRobot is still doing its best to make the Roomba better than ever, Jones left to found his own agricultural robotics company, Harvest Automation, in 2006.

Now Jones has started his second robotics company, Franklin Robotics, which is funding its latest project through Kickstarter: Tertill is a solar-powered, weed-destroying, fully autonomous and completely self-contained robot designed for your garden. Put it out there, forget about it (mostly), and it will brutally exterminate any weeds that it can find, as long as they’re short.

The genius thing about Tertill is that it’s self-sufficient. It has one button, you push that button, and then forget about the robot while it weeds your garden every day, forever. You can talk to it via Bluetooth and get updates and statistics and whatnot, but that’s optional. With a waterproof design and batteries charged by the sun, you really can just leave the robot alone and be confident that your garden will be weed free.

While it’s true that Tertill’s method of weeding (decapitation) is not as effective at pulling weeds out completely, it doesn’t make any difference: It’s not like you’re the one doing the weeding, and the robot is perfectly happy to inefficiently weed over and over every single day, until the sun explodes.

Robotics engineer Joe Jones, who helped create the Roomba, is a co-founder of Franklin Robotics, which is launching a weed-killing robot named Tertill. Robotics engineer Joe Jones, who helped create the Roomba, is a co-founder of Franklin Robotics, which is launching a weed-killing robot named Tertill.Joe Jones

As with all robots, Tertill comes with some caveats. First, you’ll need a border of some sort around your garden (at least 5 centimeters, or about 2 inches, above the ground surface) to make sure that the robot doesn’t escape. Depending on your particular garden, this could be either be more or less annoying than installing an edge wire. Tertill differentiates weeds from not-weeds the same way that it differentiates garden from not-garden: It won’t run over anything over 5 cm in height, and will instead navigate away.

During a typical day (Boston day, we’re guessing), Tertill will spend between 60 and 90 minutes hard at work, and will then spend the rest of its time sunbathing to recharge its batteries, much like myself. If you have a garden of unusual size, you can populate it with multiple Tertills, which will coordinate their schedules (although not their coverage area) so that they’re not running at the same time. This maximizes the amount of time that there’s a moving robot in your garden, which can help scare away wildlife and attract passers-by.

Tertill is on Kickstarter until July 12, and you can pledge US $250 for one that’s scheduled to be delivered in May of 2018. We’re obligated to remind you that this is a crowdfunded project and you’re backing an idea, and not paying for a product that already exists. Jones and his colleagues, though, have a lot of experience in building affordable, reliable robots, and we’re looking forward to seeing how well Tertill performs in practice.

For more on Tertill, Roomba, and why consumer robotics is so hard, we spoke with Jones via email.

Joe Jones on…

  1. Lessons From Roomba’s Early Days
  2. From a Round to Rectangular to Round Again Robot
  3. Robot Personality? No. Navigation Strategy? Yes.
  4. Chasing Chipmunks and Other Future Features
  5. “Robots Are Deceptively Hard”

Lessons From Roomba’s Early Days

IEEE Spectrum: What did you learn from the process of bringing Roomba to market, and how did that influence your approach to Tertill?

Joe Jones: I remember working in a back room at iRobot long ago when they were located above the Twin City Plaza in Somerville [outside Boston]. Roomba needed a very low-power vacuum, and I was trying to develop one that that used no more than 3 watts. I built the test mechanism from cardboard and packing tape and used the guts of an old heat gun as my vacuum source. I suddenly thought, “What do I need iRobot for? I could do any of this at home using only my own resources.” But I came to appreciate that there’s actually a lot more to bringing a product to market than just the invention phase. Depending on the product, invention may be possible on a shoestring but then there’s testing, compliance, inventory, manufacturing, distribution, customer service, and repair. All the necessary but boring stuff in which I had no interest. So mostly my view of product development matured as a result of Roomba. We have managed to do the early development of Tertill at a pretty low cost but we will need more resources for the next phase. Our Kickstarter campaign is part of that.

Spectrum: What made you decide that now was the right time to introduce a consumer weeding robot?

Jones: Opportunity. There’s a need for the application and it can be accomplished using existing technology. Had I thought of it and been in a position to work on it, Tertill would have been possible years earlier.


From a Round to Rectangular to Round Again Robot

Spectrum: It looks like you went through lots of design iterations before you ended up with those extreme camber wheels. Can you talk about the process that you went through before you settled on the final design for Tertill?

Jones: The first autonomous version of Tertill used two-wheel drive and was nearly round. But when we tried it in the garden it had problems with slopes and ruts. I did some calculations that demonstrated the limited slope-climbing ability of any two-wheel versus four-wheel drive robot. It was clear both practically and theoretically that two-wheel drive wouldn’t be versatile enough to handle realistic gardens. So we had to move to four-wheel drive. Robots that have a coverage task (like robot vacuums or weeding robots) must come into contact with all obstacles in their workspace. (If you don’t touch the obstacles then you leave an unprocessed ring around each one.) We made Roomba round with the two wheels on a diameter because we knew that no matter how complex the geometry of nearby obstacles, the robot could always spin in place to find an escape path. That strategy simplifies motion planning (ergo the almost round shape of gen 1) but complicates the mechanics—it’s easier to fit components into a rectangular robot.

Franklin Robotics' Tertill weed-killing robot prototype 0 to prototype 3From left: First prototype and generations 1, 2, and 3 of the robot, which went from two-wheel to four-wheel drive to improve its ability to drive over slopes and ruts.Franklin Robotics

Now that we had to have 4WD, a rectangular shape became much more attractive. This brought us to generation two—a long way from round but narrow enough to thread its way between closely planted rows. Our hope was that the garden environment was forgiving enough that we could get away with a non-round robot. Traction was great, the robot could go up and down slopes with no problem but overall maneuverability was poor. Trying to turn corners Tertill swept out a wide area and sometime swung sideways into plants. If we had complete information about the local environment we could plan clever paths to avoid such problems. Unfortunately, complete information requires fancy sensors. Fancy sensors cost real money and result in robots that customers are unwilling to pay for. If we couldn’t afford the sensor we had to change the robot. In generation three we tried a compromise keeping the 4WD but making the robot a little closer to round. Better but no cigar. The robot could still get hung up in a corner that wouldn’t have given Roomba the slightest problem. Eventually, Tertill could usually extricate itself from such spaces but the many attempts required for final success was problematic. Every failed motion can cause the wheels to dig in farther, changing the garden surface and making escape more difficult.

Franklin Robotics' Tertill weed-killing robot prototype 4 to final versionFrom left: Generations 4, 5, and 6, and the final version (available on Kickstarter) of the Tertill robot, featuring a round shape and camber wheels.Franklin Robotics

So the robot had to be round. That brought us to generation four. Packing four wheels into a round robot is tough. The wheels need to be as large as possible to get over rocks and holes and you want as large a wheelbase as possible in both dimensions to minimize the likelihood of the robot tipping over. Generation four worked better than all previous robots but the wheelbase was narrow and we began to worry about high-centering. When something the robot has driven over (e.g. rock, dirt clod) touches the underside of the robot, that object supports some of the robot’s weight. The wheels thus lose traction and the robot may become stuck. Thinking about this and the wheelbase issue led us to the extreme camber wheels. This new configuration moves the contact point of the wheels out to the edge of the shell and means that more of the underside area of the robot is movable. That is, there’s less area that can participate in a high centering event. And we get the added bonus that we have more room for the whacker—we can cut a wider swath. We think the high camber idea is interesting enough that we made it part of our patent (pending).


Robot Personality? No. Navigation Strategy? Yes.

Spectrum: Does Tertill have a personality? Do you find that users develop emotional connections with Tertill like people commonly do with Roombas?

The most common response upon seeing Tertill for the first time is probably, “That’s adorable!” So I expect many Tertills will be named and become part of the family.

Jones: Robot personality is in the eye of the beholder, in my view. Personality is not part of Tertill’s specs. And those of us who know it best (hardnosed technologists all) have not developed emotional connections. But the little robot is very engaging. The most common response upon seeing Tertill for the first time is probably, “That’s adorable!” So I expect many Tertills will be named and become part of the family.

Spectrum: iRobot always made a point of telling us that Roombas are making careful, data-driven and sensor-based decisions about where to go, as opposed to just bouncing randomly around a room (which is how it appears at times). What kind of navigation strategy does Tertill use?

Jones: I agree that from a high level it looks like both Roomba and Tertill just bounce around randomly. But both contain some hidden subtlety that makes a huge difference in performance. The effects have to do with coverage and escaping hazards. A robot that does nothing but bounce randomly has a hard time fully covering a cluttered space. The trick is that some fraction of the time, rather than bounce, the robot should follow an obstacle or a wall or a row or crops. This allows it to escape tight spaces and fully cover a cluttered area. (If you’re really interested, the method is described in my book, “Robot Programming: a Practical Guide to Behavior-based Robotics.”) The second issue is more complex; it arises any time the robot must respond to a second hazard while still dealing with a first—say encountering a wall while avoiding a cliff or turning away from a plant and encountering a rut at the same time. The earliest version of Roomba (which didn’t deal with simultaneous hazards) had maybe half a dozen behaviors. But by the time we had worked out the bugs of all the special cases, Roomba had over 50 different behaviors. We on the Roomba team used to feel quite smug when we encountered a knockoff robot where the developers had clearly stopped working after writing just the first six behaviors. Such robots got stuck frequently and had a hard time getting away from even the most benign hazards.


Chasing Chipmunks and Other Future Features

Spectrum: Are there features or capabilities that you originally wanted Tertill to have, but had to cut due to cost, reliability, or practicality?

Tertill can move suddenly at random intervals, and nearby pests are likely to be frightened away when this happens. Later models may include motion sensors enabling Tertill to give chase.

Jones: Yes, we had hoped to include deliberate pest scaring and extensive monitoring of the local environment in the first iteration. Unfortunately, our initial approaches to those features proved to be too complex, too costly, and required too much development time. We did, however, find lower impact ways to achieve part of our goals. Although the robot doesn’t detect the presence of pests (like rabbits and chipmunks) we can schedule Tertill to move suddenly at random intervals. Any nearby pests are likely to be frightened away when this happens. Later models may include motion sensors enabling Tertill to give chase. We would like to add fixed sensors in the garden to monitor things like soil moisture, conductivity, and so on. A future robot will collect data from these sensors and compile a report alerting the gardener to any developing problems that may affect plant health or yield. The initial version of the robot will collect some of the information we’d like to provide. This includes insolation, local temperature, how often it finds weeds to chop.

Spectrum: My Roomba works best when I keep the floor mostly free of clutter and electrical cords. Is there a way to optimize a garden for Tertill?

Jones: Slopes that exceed 40 percent grade, big rocks, big ruts or holes, and persistent puddles all make it more difficult for Tertill to operate. Gardeners will make Tertill more effective if they minimize or eliminate those features. We choose Tertill’s diameter after surveying the sort of plants that are most often grown in home gardens. My research suggests that most plant rows should be more than 8 inches apart and within a row most plants grow best if spaced at a minimum of 8 inches. Respecting the recommended minimum plant spacing (or even giving them a bit of extra room) is good for both Tertill and your plants.


“Robots Are Deceptively Hard”

Spectrum: Why is it so difficult to build reliable, affordable consumer robots?

Jones: Robots are deceptively hard. My co-founder at Harvest, Clara Vu, likes to say, “A robot application that a technology grad student thinks they can easily do in a couple of weeks has an outside chance of actually being practical. Anything harder is impossible.” I think the issue that misleads developers is that robots have different strengths and weaknesses compared to people. What that means it that any application you want to roboticize must be re-imagined from the ground up. With Tertill we didn’t just take the base of an RC car, strap on a weed whacker, and call it done. That would work fine if you or I were at the controls. We would rarely drive the car into any debris or terrain from which it couldn’t escape. But the robot doesn’t have our sensory and cognitive advantages, it will simply drive itself into any rut or mass of foliage and only realize that there is a problem when it stops moving. Thus we spent a year designing and redesigning the chassis and mobility system to make it possible of the robot blunder into difficult situations and still manage to get out.

Spectrum: What are you most excited about in robotics right now?

Jones: I’m always interested in low-cost robots. Robots that just about anyone can afford have the greatest potential to create new industries and change the world. Currently I’m thinking that cameras connected to deep learning networks is the technology with the most promise to enable a new wave of low-cost robotic applications. It’s possible that I only think this way because I don’t yet know a lot about deep learning and am blissfully ignorant of its shortcomings. But I’m trying to learn more.


Updated 4:46 p.m. ET

The Conversation (0)