January 19, 2017
Ben Chostner of Blue River Technology explains why and how robotics represents the future of farming.
Robots aren’t merely the domain of auto manufacturers and high-tech warehouses. Ben Chostner, head of business development for Blue River Technology, is helping to bring robotics to a place you might not expect to see it: the farm. Outside the controlled environment of a factory, what can smart farming equipment do? Chostner spoke to us from his Silicon Valley offices, where Blue River employs roughly 50 people, working toward its goal of improving farming efficiency while reducing agricultural waste.
PwC: How are robots being used in farming?
Ben Chostner: Every industry, including agriculture, has some jobs that are non-differentiating and onerous. And in some cases, such tasks go undone because either no one wants to do them or the business cannot justify the cost of doing them. Skipping these tasks limits the performance the business is able to achieve.
Our vision is that smart agriculture needs smarter equipment. We are developing a smart sprayer that we call See and Spray. These machines move around the field, see every plant, and know whether it’s a crop or a weed. If it’s a crop, the machine knows important characteristics about the plant, like its health, size, and maturity. We use all that information to make very fast, very smart application decisions: apply herbicide just to the weed, or fertilizer or fungicide just to the crop plants that need it, and so forth.
PwC: How does your technology work?
Ben Chostner: The technology has four key components. The first is a ruggedized seed line box—a ruggedized electronics enclosure that includes all of the intelligence of the system, everything that’s needed to see and spray for an individual seed line. You can imagine the challenge. We’re taking out of the server rooms and the [computing] cloud some electronics and hardware that do really advanced machine learning and deep learning. And we’re bringing them to an agricultural environment where they’re exposed to vibration, dust, and weather conditions.
The second component is the camera in the front of the box. This camera takes pictures of the field the device is moving over. A deep learning algorithm identifies all of the crops or weeds in the field and then feeds that information into the sprayer system. The third component is the sprayer. The sprayer can deliver the spray via a nozzle very precisely, in 1-square-inch patches, consistently and in real time. The final component is a second camera in the back of the device. This camera observes where the spray went relative to where it was intended to go.
Together, these components allow us to have a closed loop system, where the system monitors its own performance and adjusts itself on the fly. The farm environment is harsh and the spray nozzle may get hit by a clump of clay and go off course. We can reorient the nozzle before the next spray if the previous spray was deemed to be off target.
PwC: You chose an architecture in which your computer is onboard the device, and all the processing happens locally. Is there any communication between units? If you have four running in tandem, do they talk to one another?
Ben Chostner: Generally speaking, it’s not needed. To be fully modular, the machines should be independent, but we pass some information between them and we send some information back to the cloud to our service center for support purposes. Sometimes the farmers care about what’s happening in seed lines that are next to each other. But that communication happens at a low frequency.
PwC: What is the significance of a 1-square-inch resolution? How was that reached?
Ben Chostner: Weed control is all about selectivity. It’s very easy to eliminate weeds; it’s much more difficult to eliminate weeds but leave your crop intact. The history of weed control has been all about this search for selectivity. For thousands of years it involved hand tools: people used their own eyes and hands, identifying weeds and taking them out. Then tractors allowed cultivation between the rows at speed and on a bigger scale. Chemistry developed herbicides about 50 years ago, so you could selectively control some plants but not others. About 20 years ago came genetically modified plants; the crops were modified to be resistant to the herbicide, and you could apply it everywhere.
The future of digital agriculture is to use smart machines to regain selectivity. You can use any control mechanism you want, any chemistry, any method, and apply it just to the weeds, only where you want it. You don’t need to worry about whether that control method actually kills your crop, because you’re not touching your crop. And so a 1-inch resolution allows you to apply that in the appropriate places without damaging the crop.
PwC: How does machine learning function in your device?
Ben Chostner: The beautiful thing about machine learning is that the more you use it—the more situations you expose the algorithm to—the better it gets. The solution we deliver to the farmer will already have a trained algorithm in it. That does not mean we do not learn from day-to-day usage. From every field that we go into, we save a few images and send them to our learning database, and we’ll retrain our algorithms. So maybe in that field that day, it’s not getting the benefit of that new learning. But tomorrow when it runs, or when a machine in another field runs, it will gain the benefit of that continuous learning.
PwC: If a farmer got the system for lettuce and now they want to grow corn, what is required to adapt the system to a new kind of plant?
Ben Chostner: Mostly it’s collecting images and building up the training database. It’s similar to the way many photo-sharing services work. The first time you upload a photo of someone, the service has no idea who it is. But by the tenth time you upload a photo, the service knows exactly who your spouse or child or friend is. The photo-sharing service doesn’t rewrite the algorithm for every new person who shows up. It just trains the existing algorithm. It’s a similar concept for us.
We have also built some flexibility and configurability into our technology. We can take one of these seed line boxes and customize it to different applications by changing the software and changing the toolbar. So we can ask a farmer what crop he has and in what configuration he is planting it—spacing between rows and how many rows are planted at a time—and deliver a custom configured spray machine to any farmer around the world.
PwC: How have your farming customers responded to your product?
Ben Chostner: Solutions such as ours provide quick feedback to the business. A day or two after our machine does a demo in a customer’s field, they can walk through the field and give us a thumbs-up or a thumbs-down. Did we control the weeds or not? They don’t have to wait multiple seasons for yield or profitability data. They can just go through and say, “Yes, this system controlled the weeds.” If the costs fit with what is expected, they would be in a position to make a decision.
There are some things that customers need to do, such as change their growing practices. Over time, we’ve been able to understand what those are to help them get the most out of this equipment.
PwC: How is your system affected by some of the evolving data and interoperability standards?
Ben Chostner: These are early days from a data standards perspective. Our data set includes every single plant in the field: every crop, every weed. All of the information is plant by plant. That level of resolution and precision isn’t really imagined in some of the data standards that are being written right now. From an equipment connection perspective, there are standards in cars that we’re looking into for displays or for interfacing with tractors. But because our system has so much information to pass around, that protocol doesn’t have enough bandwidth for us. The standards being worked on today don’t quite apply for our equipment.
PwC: How do you think farmers will get value from such large data sets?
Ben Chostner: A couple of farmers have asked us for a plant-by-plant data set of their field. And we give them a spreadsheet with 3 million entries. They realize quickly that they can’t do anything with that, and we both look at each other and say, “Well, gee, this isn’t useful.” The challenge is to figure out how to turn some of that really dense data into an actionable decision or an intelligent conclusion. We’re working through that, and I think there’s a lot of potential. Much of the benefit of the data we collect is in those real-time decisions—knowing whether something is a crop or a weed and making a smart application decision. We have an eye toward taking more advantage of some of that data in the future.