Building Autonomous Greenhouses with AI and Robotics

Gilwoo Lee is the founder and CEO of Zordi, a company that builds and operates autonomous greenhouses with mobile robots and AI. They raised $20M in their most recent funding round led by Khosla Ventures. She has degrees from MIT, CMU, and University of Washington.

Gilwoo's favorite book: Masayoshi Son's 300-Year Plan (Author: Takashi Sugimoto)

(00:01) Introduction
(00:08) What is a Greenhouse?
(01:08) Greenhouse Usage in Different Regions
(02:45) Advantages and Challenges of Greenhouses
(06:15) Crops Suitable for Greenhouses
(08:23) Introduction to Autonomous Greenhouses
(10:40) Technology Stack of Autonomous Greenhouses
(12:17) Factors Affecting Crop Growth and Automation
(14:00) Data Requirements for AI Models in Greenhouses
(15:14) Upgrading Traditional Greenhouses to Autonomous
(17:26) Tasks Managed by Robots in Greenhouses
(18:45) Engineering Challenges in Robotics for Greenhouses
(20:16) Future of Robotics in Greenhouses
(21:39) Robotics Supply Chain and Standardization
(23:23) Verticalized vs. General-Purpose Robots
(27:16) Robotics Foundation Models
(29:52) Economics of Autonomous Greenhouses
(33:16) Future of Autonomous Farming
(34:01) Rapid Fire Round

--------
Where to find Prateek Joshi:
 
Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.602)
Gilbu, thank you so much for joining me today.

Gilwoo Lee (00:05.173)
Thank you for having me, nice to meet you.

Prateek Joshi (00:08.276)
Let's start with the basics. What is a greenhouse?

Gilwoo Lee (00:15.047)
Yeah, great question. So Greenhouse is essentially...

a structure, physical structure that has either polyfilum or glass or some sort of a material that lets the sunlight go through while protecting the inside from rain or any extreme weather conditions that are outside. And there are many different varieties of greenhouses going from very modular things like hoop houses

single hoop house could be as small as 3,000 square feet to really big glass greenhouses, steel, almost like steel building type of structure, be as large as 20, 30 acres.

Prateek Joshi (01:08.11)
And in our day-to-day lives, obviously greenhouses have been used in practice for a while. So can you talk about what percentage or some metric, just to give a sense of the scale of how often do we see it in our day-to-day lives?

Gilwoo Lee (01:26.443)
Yes, in the US, only about 1 % of the crop land is recognized in the USDA Act census as greenhouses. So it's not super common for you to drive around and expect to see greenhouses unless you're in a very agricultural town. However,

even like a very small thing that you can put in your backyard that just like a tiny thing also counts as a greenhouse. So in that sense, actually, I think a lot of people just have a small thing that they grow in in green in their backyard. And then you'll see them a lot. But it you would see them a lot more in other countries, Canada, Mexico, UK, Spain.

Korea, Japan, just to give you an idea about 40 % of cropland in Korea, which is where I'm from, is greenhouse type. So you see them a lot as you drive and if you go to the Netherlands about 99 % of the cropland is actually greenhouse.

Prateek Joshi (02:45.048)
What are the challenges? Meaning actually both questions. Then one, why bother doing greenhouses? Like why not just regular farming? So what are the advantages? And also if you want to do it, what are the challenges?

Gilwoo Lee (03:02.645)
Yeah, great question. So.

The biggest advantage of greenhouse is that it is climate resilient. It's not completely agnostic to climate, but you can do a lot of things in addition to just putting a structure to protect.

the plants, assuming you're growing plants inside, from cold weather or even from humid weather, things like that. So here's one interesting story. Literally, if you type and look up in the Wikipedia the history of greenhouses, I was personally surprised to find this. South Korea in the old, it is the first country to...

have started or have a written record of a version of greenhouse hundreds of years ago. And the reason being the royal dynasty, dynasties, the king wanted to have citrus during the winter time. And because that doesn't actually work naturally, they ended up putting, inventing something to make that happen. So the biggest, that kind of comes back to your question, what is the biggest benefit?

It is climate resilient. You can extend the season. You can do even with modern technology, completely flip the season. It gives plants the basic protection and a lot more in addition to that, depending on the infrastructure you put in to be able to completely flip the season if needed. compared to traditional agriculture that also reveals some of the challenges.

Gilwoo Lee (04:53.693)
it does require some capex, significantly certainly more than just doing everything in the field. So when in a country with a lot of good climate, weather, land, as in California or Florida, when there's like you have abundant land with good climate, you don't have to really start building greenhouses. So that's why in the US, greenhouse wasn't

the number one choice for doing agriculture while in these other countries where they have more extreme weather conditions, it started a little bit earlier. Now, coming back to where we are now, because of extreme weather that's happening more frequently these days, and it doesn't necessarily mean hurricane or wildfire. For growing crops, things as sensitive as more rain or just too dry weather, all of those

are still some version of extreme weather that stresses out plant because agriculture gets so heavily impacted by those events, more and more greenhouses of all sorts of kinds are actually being put up even as we speak in the US a lot more than it used to.

Prateek Joshi (06:15.766)
And if you look at all the crops that can be grown in a greenhouse, so what crops are more suitable on average for a greenhouse versus some crops that are not? Or maybe more importantly, how do you decide, okay, this crop might benefit from a greenhouse versus that crop? You may as well just grow it in the

Gilwoo Lee (06:37.973)
Yes, so great question. And I think there were a lot of misunderstandings about what should be grown in greenhouses or even to an extreme versions of different type of indoor farms over the past 10, 20 years that made this industry pretty challenging or had a lot of failures. Long story short.

It has to be the answer has to come from the basic unit economics. What does it economically make sense to grow in greenhouses? And often, most the immediately obvious things is you wouldn't benefit from growing corns or soybeans or rice.

in greenhouses you can and you'll certainly do slightly better than doing it in the fields because of all the benefits but the cost benefit is just not there while you would typically benefit from a growing things like leafy greens high value what we would often call as high value fruits and vegetables as tomatoes cucumbers bell peppers strawberries

These are very common examples and more and more of those high value crops are being added to greenhouse categories as we speak because there's more research, ongoing research in horticulture to make it easier and more beneficial to do that inside of greenhouses.

Prateek Joshi (08:23.542)
Now let's move the discussion to what you're building, which is autonomous greenhouses. Now can you explain what an autonomous greenhouse is and how does it differ from a traditional greenhouse?

Gilwoo Lee (08:38.549)
Yeah, there's the more conventional greenhouse. I'm going to make maybe three distinctions here. There's like very traditional old school greenhouses that are more like hoop houses or very small, easy to build ones with minimal, almost zero infrastructure. Like there's no heating systems, there's no circulation fans, just protection from the rain.

There's also then more of modernized greenhouses where you put all of those HVAC systems and sometimes you put in LEDs, sometimes you put in misting systems and these get automatically controlled through just like how you control the heaters in your house or you can set the timer. So that's a version of autonomous greenhouse that people are very familiar with.

And it's been around for at least 10, 20 years and they're kind of gradually changing as these technologies have developed. What we are building is a, on top of that, AI robotic greenhouses. And what I mean by that is actually, even with that level of the pre-bought, I just described as autonomous greenhouses, there's a ton of horticultural expertise in manual labor that has to go in.

to grow these crop right and then harvest and pack and deliver. And we are with our AI robots doing that, making all the horticultural expertise and manual work kind of being taken care of mostly by our robotic agents and intelligent agents. And that makes it a lot easier to replicate with minimal mistakes, minimal risks with a lot less of the

almost non-existent labor.

Prateek Joshi (10:40.096)
And if you look at the technology stack of an autonomous greenhouse, can you talk about all the technologies that enable autonomy in the greenhouses today?

Gilwoo Lee (10:55.551)
Yeah, so in the conventional sense of autonomous greenhouse, that's often being referred to, it's the basic things that needs to be autonomously running are irrigation control and temperature control, humidity control, which all come with ventilation, could come with ventilation.

or schedule or time based schedulers, things like that. When there's a level of autonomy that could be done by basic scheduling. But these days, the industry is moving towards monitoring, connecting the different sensors and monitoring the more specifically what is the water temperature, what is the soil temperature, and then tying that to these controls, environmental controls to make

better decisions.

Prateek Joshi (11:55.374)
And if you look at all the factors that impact crop growth, are all those factors being controlled or driven by an autonomous system? Or are some factors more likely to get automated versus some factors where you just can't do much?

Gilwoo Lee (12:17.407)
Without our AI and robots, the things that you can control autonomously is fairly limited. Or fairly limited in the sense that it's about sort of like, yes, you can control your temperature, just like how you change your thermostat. You can control.

the irrigation schedule, how many times you water, those are fairly easy to automate and have been already automated. Without our AI robotics, it is almost impossible to have a real intelligence, unless you have a really good farmer slash grower on site making those intelligent decisions to do, should we change fertilizer to achieve certain

different profile, flavor profile, or improve the yield, or should we prune certain area because this is getting the plants are being overgrown. Those things require horticultural expertise and without our, that's what our our proprietary technology brings in. Without that, you really need a very, very skilled set of experts to make the right decisions.

Prateek Joshi (13:39.25)
And to build this AI model, obviously you need data. So can you talk about all the data that goes into building a model? And also as you deploy this model on the field in a real greenhouse, what data comes back to you as part of this process?

Gilwoo Lee (14:00.149)
Yeah, the type of data that is needed is almost like trying to identify, yeah, could be very broad, as in might you want to collect microclimates in the greenhouses, certain regions are more warm than other regions, or you could you also want to capture the growth of different plants.

You know, you're just like human beings, they were organic, so they grow in certain variants. And if the plants in hotter regions are bound to grow in slightly different pace than those in colder regions, they might come require different attention. So those are the things that get captured. by growth, I mean the height of the plants, the number of leaves they're pushing, how big those leaves are, number of fruits they're pushing. All of those need to be captured.

So those are being collected by our robotic agents and the sensors and then being coming back or being actively used to build the models that we make.

Prateek Joshi (15:14.094)
Right. And if you look at all the traditional greenhouses today, can every single one of them be upgraded to an autonomous greenhouse? Or are there some greenhouses that are way more likely to realize a big benefit or a big jump in ROI by upgrading to autonomous?

Gilwoo Lee (15:37.333)
Yeah, that's a great question. Technically, I would say...

Yes, all of them can be upgraded to become more autonomous. And I say that with confidence because autonomy comes with different levels of autonomy. without going to the full extreme of putting robotic agents and letting the AI agents make decisions, you as a farmer would still benefit a lot from automating most some of the core tasks.

that said I would say greenhouses that are growing certain type of crops like fruiting crops would benefit most from upgrading to the

I would say frontier version of the greenhouses that we are pushing forward with robots and AI because those are the ones that need very constant attention and you can really maximize the benefit or the yield and quality of your crop by putting that level of effort.

Prateek Joshi (16:56.366)
Now you mentioned the tasks. So I think that's a good point to talk about the robots. So obviously when the AI model, it looks at the real time data, makes a decision. And to put that decision into action, you need actuation, which is a robot which comes in and takes the action. can you talk about what are all the tasks that the robots manage in this environment?

Gilwoo Lee (17:26.695)
Our end goal is really to be able to have, even with multiple robots, have all the tasks be managed and done by robots. We obviously, for a good reason, start with the ones that require, A, most expertise, horticulture expertise, and B, the biggest

components of manual labor. For example, with those two categories, when it comes to manual labor across these fruiting crops, would be harvesting, sorting, and packing of the fruits. That would be the first task. And then if you break it down, go down the list on the manual labor further, pruning leaves or removing smaller fruits or removing smaller flowers or removing bad leaves.

not just like old leaves, those all kind of come under the category, or depending on if you're an organic farm and your farm relies on beneficial insects to control pest pressure, that will about releasing beneficial insects. All of those come to that category of like manual labor. And we really tackle those by combining economic impact that we can bring and how hard or feasible it is given the latest technology.

Prateek Joshi (18:45.676)
What are the engineering challenges you face when you have to integrate robotics into such an intricate ecosystem?

Gilwoo Lee (18:54.677)
Yeah, I think the interesting challenges are that it's not, it is literally organic, as in every single plant is somewhat different. And I would say our challenge conceptually is not different from trying to automate how doctors and nurses make decisions about their patients and how to follow up with treatments when it comes to decision making.

except we're dealing with much easier version of those. It is where we do benefit from the fact that we can still make some mistakes. We can still crush some fruits and that's still okay. But the challenges are just naturally the fact that every single kind of branch of fruits look somewhat different in.

So the robot has to be pretty good at dexterous fine manipulation as well as being able to detect and sense different parts of plants.

Prateek Joshi (20:01.602)
And what kind of future developments, technological developments in robotics could enhance the functionality of autonomous green house? Like what do you wish the robotics community or robotics builders do next?

Gilwoo Lee (20:16.385)
Yeah, we are the robotics builder in this space, so we get to choose what we do. Honestly, I think.

Prateek Joshi (20:19.946)
Right, right, yes, right.

Gilwoo Lee (20:29.853)
Not as a robotics builder necessarily. The simulation side of things getting more more real, not just the physics engine part of it, but more of the perception of rendering part of simulations. I do think that as we, the robotics community, start

playing with it a lot more and I know it is happening as we speak and figure out how to integrate that to develop the next robots or next robotic enable robots to do more complex tasks with minimal real world experiments because that's the really costly one. I think that will accelerate not just more robotic greenhouses but also just broadly speaking robotic applications.

Prateek Joshi (21:25.71)
And as a builder of robots, is there anything on the supply chain or manufacturing side that you wish would be better for you as a builder of robots?

Gilwoo Lee (21:39.569)
that's a great question. I do find it exciting that robots are getting cheaper and cheaper and more accessible versions of robots of all sorts, even including humanoids, are coming out every year across different continents, different countries.

So I wouldn't say that there's specific challenges currently in the supply chain that I think the community needs to resolve. One thing I think we as a robotics industry, especially just me coming out of academia have found it much harder to achieve relative compared to other communities, like computer vision community is having standard set of problems to

solve and compare results, which is fair because robots are all different and the problems are there's just real world problems are so hard to standardize, right? So that I think the community as a whole will certainly benefit from having some standardized way to build data sets and use specific type of robots and compare different algorithms and there have been

many, many approaches or proposals to do that. So it's not to say that people are not working on it, but I do think that still I think as a community that we are trying to achieve for a better, faster acceleration of more robotic developments.

Prateek Joshi (23:23.436)
That's actually a good segue into my next question about the form factor. Obviously, robots of all shapes and forms, people are working on it, they're building on it. And on one end, we have the very humanoid robots. They can be deployed in a variety of situations, but they are generic, versus a verticalized robot where, for example, for forming, you can really tailor the robot to the needs.

of an autonomous greenhouse and you don't have to worry about whether the robot knows how to cook, for example, or how to talk on a phone. So between verticalized robots versus humanoid robots with broad applicability, where do see the future going? Is it one or the other? Is it a combo of the two? What do you think is going to happen here?

Gilwoo Lee (24:12.031)
Yeah, that's a great question that I know there's a lot of heated debates and I share both very bullish optimism on the path on generalizable robots and like humanoids and also on the value of very practical, almost tailored application specific robots. And I do think that

robotics companies or just robots in general need to demonstrate the value to the world soon. And there is, it is a lot easier to achieve that with these robots that are application specific. I think,

those have to come first or those will come first when it comes to kind of practically delivering value to big customers or small and large customers and then eventually humanoid and more generalized robots, they will have to find applications where they will be providing more value than these very application specific robots. And I'm sure

It'll happen. It's more so I do think that I genuinely believe that it has to be bi-directional. If I were to be a little bit more specific, though, I do think that the application specific ones will certainly show more results faster just because of the nature of it.

Prateek Joshi (25:51.18)
Right, and also for a given use case in any vertical, wouldn't the application specific robots be like 10x faster and 10x cheaper because they don't have to worry about 99 % of other use cases and edge cases. So just by that, basic physics of it would make it faster and cheaper for that. So customers in that vertical are more likely to choose a domain specific robot, right?

Gilwoo Lee (26:11.869)
Exactly. Yes. Yes, that is very true. Where it will start changing is that there's always when you're thinking about just autonomous driving or autonomous greenhouses, there's long tail end of set of tasks that you want to automate and you don't necessarily benefit or cost wise even from automating building.

those task specific things over and over again at the end of the tail end. And then if by then, by the time these application specific robots have filled in like say 70, 80 % of the tasks that need to be done, if a generalizable robots, even at a much slower pace, because they're general can tackle those problems, then I think that alone would bring

huge value even if they don't end up wiping out the custom application specific robots.

Prateek Joshi (27:16.206)
Yeah, that leads me to the next question on robotics foundation models and people are working on it. if you look at all the robots, there are certain things that are common, like sensing, analyzing, processing, actuation, adhering to physics. So some of the things that no matter what robot you build, these things will remain more or less common. So a robotics foundation model, one, what are the things that foundation model needs to do well?

And also what are the things across all the use cases that are common enough that a foundation model can actually be practical in the real world.

Gilwoo Lee (27:57.473)
Honestly, I feel a little bit hesitant to answer. Those are very, very critical questions that I do think that what is hard about robotics, foundational models is just that definition of those things. As in, it is...

Because of the nature of the set of tasks and the nature of the things that we want robots to do, think people are really trying to figure out what that should look like. And I do think that right now, the community think there's, everybody has somewhat different definition of what that should be. And I think some, so long story, long winded way of saying, I don't think I can provide confidently a set of like answers to that.

But I do think that whoever can define a generalizable set of tasks or ways to transfer skills from one to another or build a system where the different type of tasks can be picked up quickly.

where those specific set of tasks end up being widely applicable. And I say very broadly because I actually do think that even though I say widely, the value of house cleaning robots versus the value of firefighting robots, they're very, different. So ultimately, the set of tasks that these foundational models can immediately solve, I think, will have to be very economics driven.

Prateek Joshi (29:32.814)
Alright, bye.

Gilwoo Lee (29:44.538)
And as long as you can cover that, those tasks, I think it is still foundational enough to be a breakthrough.

Prateek Joshi (29:52.566)
Right. That's a great question. on the topic of economics of a greenhouse, coming to the more practical side of it in the sense that if you go to a farmer, obviously all the research is great, but it has to make sense financially. what are the economic factors involved in making a decision on using autonomous greenhouses from the farmer's perspective? And

Is the investment scale accessible to most farmers? Like how big or how accessible is it?

Gilwoo Lee (30:28.821)
Yeah, that actually, those are the questions that we have been trying to build our systems and the business models around from day one in the sense that ultimately.

it has to be huge return on investment. And B, so that's number one. And what I mean by that is fast payback period or literally just margin coming out of using autonomous greenhouses with robots versus doing things in the field. Secondly, the investment has to be within

You know, it has to be affordable and it has to be something that's A, even just total sum of it is not too expensive and it is, and B, it has to be something that's most of it is easily financeable. So what we are providing or our systems

Our systems are designed to work well, even with very old school traditional hoop houses, as well as glass greenhouses, which means that the CAPEX investments that farmers need to make is very minimal.

And certainly if they're more willing to invest more in like expensive greenhouses, that is certainly valuable to them. But that's not that's not the barrier that is that they need to overcome. And our robots are essentially much cheaper than.

Gilwoo Lee (32:11.965)
the labor that they need to get and also a lot more consistent and typically for farmers, this is just true in any modern countries in many cases. It's not about, am I displacing labor with the robots? No, typically they just can't find the labor. So the fact that we are able to fill in the gap is huge. So that ends up being very quick return on investment combining those two.

And luckily last point is that there are a lot of great ways for them to finance these upgrades through USDA or farm credit like institutions because these are just legitimate value add to their farmland.

Prateek Joshi (32:58.882)
I have one final question before we go to the rapid fire round. this is about going beyond autonomous greenhouses to the broader topic of autonomous farming in general. Now, how do you see the future of autonomous farming evolving in the next five years?

Gilwoo Lee (33:16.799)
Yeah, I actually have a great friend that I have deep respect.

his name is Kenny Lee and he runs a robotics ag robot startup called Eigen and his vision that I deeply share is every farm should have a robot and my vision is that every farm should have obviously a lot more than one one robot they should have intelligent robots actually running the farms so I do think that in the next five years you're going to see a ton of that it'll just be just like how drones are so commonly being used in

like media and everywhere, like some version of robots will be so cheap to get that it's already a no-brainer. They're just like entering the market as we speak.

Prateek Joshi (34:01.39)
Amazing. And yes, I share that vision too. I think it's a very strong use case for farming robots. So just do like take off all the basics and tasks that should be done by machine. So with that, we're at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready? All right. Question number one, what's your favorite book?

Gilwoo Lee (34:21.3)
Yeah.

Gilwoo Lee (34:26.401)
It's a book about Masashi Son. I don't think there's an English name. It's about, the title is Ambition, a vision of Masashi's 300 years worth of empire.

Prateek Joshi (34:41.774)
Amazing, that's actually I have not read that one yet. I'll pick it up. All right, next question. What has been an important but overlooked AI trend in the last 12 months?

Gilwoo Lee (34:54.175)
I don't, I thought about it, I couldn't answer that question. The trend is so rapidly changing. So I think it's best for me to do for that.

Prateek Joshi (35:01.678)
Next question. What's the one thing about autonomous greenhouses that most people don't get?

Gilwoo Lee (35:10.655)
The one thing that people don't get is the complexity of crop care decisions that needs to be made every day.

Prateek Joshi (35:24.458)
What separates great AI products from the merely good ones?

Gilwoo Lee (35:30.324)
Mmm.

usability of making it actually effective versus just being a random chatbot. I think you're already seeing it everywhere, whether I'm making a very specific example, but whatever AI application it is.

Prateek Joshi (35:47.232)
What have you changed your mind on recently?

Gilwoo Lee (35:52.863)
Recently, My recent change has been, the pace of simulation. How fast it is becoming better and better, especially with generative AI.

Prateek Joshi (36:07.628)
Yeah, that's actually, I love that one. It's I'm keeping a close eye and I think it's going to unlock many good things. Next question. What's your wildest AI prediction for the next 12 months?

Gilwoo Lee (36:20.978)
I don't think it counts as wildest, but I do think that you are going to see a lot of complex decision making done with AI agents in the physical space.

Prateek Joshi (36:32.879)
Again, I agree with that. I love bringing AI and software, intelligent software in general to physical and for us. So it's so much to be done here. It's very exciting. All right, final question. What's your number on advice to founders who are starting out today?

Gilwoo Lee (36:49.845)
Be very bold. I think the market has been so hard that it is hard to be bold if you were to start it right now. But that is the only reason for you to start any business, any startup, be very bold in what you want to build.

Prateek Joshi (37:08.046)
amazing. Gilu, this has been a brilliant episode. love the first of all, I just love AI and physical infra. That's the combination is very potent. It's right for disruption and you're building something amazing. So, so glad you're able to appear on the podcast and share your wisdom.

Gilwoo Lee (37:26.325)
Yeah, well, thank you so much. It was really fun and looking forward to both the my episode and the feature ones. Thank you.