Robotics vs Automation in Cannabis Manufacturing ft. Nohtal Partansky

Listen Now

Nohtal Partansky: It's kind of continuously modifying its own performance based on the feedback that it's receiving from the situation. Bryan Fields: Nohtal, cannabis facilities say they're "fully automated" versus "automated." What is the difference? Nohtal Partansky: Yeah, so fully automated would mean the end-to-end process of whatever you're talking about is automated, meaning you basically have someone supervising the process as opposed to interacting with it. We were just joking about an automated label maker — you can have an automated label maker, but that doesn't mean you have an automated label process. A label process is many times printing variable data onto a label, applying that label onto the part, and sometimes applying multiple labels. If you just have a label maker, that prints labels, but then you need something to feed it into a system that applies the label, and that's only one part of a probably much bigger process. When I ran my co-packing company, we had a mostly automated packaging line for jars, which I think is probably the only piece of cannabis manufacturing that can really get to like 90% automated. You'd load flower into a hopper, and that hopper would automatically dispense it into a line, and that line would make what are called weighments — little weighed sections of material. Then it would combine those weighments into an eighth and drop them into a jar. That jar would index around and go to a capper, which would spin a cap onto it. Then it would go to a labeler, which would take a reel of labels, laser-print the variable data onto that label, peel it off, and slap it onto the jar. Then it would go to another section and peel off another label — a lollipop tamper-seal label — and kick it out. So you had human touch points on the loading of the flower, and another touch point where people loaded jars onto an accumulation table that then got indexed out. That's pretty close to fully automated. The last piece we had for some applications was an arm that would pick up the jar and put it into a box — we didn't have anything to close the box, that was done by a person. I wouldn't even call that fully automated, because if you're fully automated, aside from loading and supervision, everything is done by the machine. But if we'd had something to close the box too, it probably would have been fully automated. In that scenario, instead of five or six people running the line, you've got two people guiding the machine, ensuring continued success, and allowing it to be a lot more consistent with production numbers and process. Bryan Fields: Is it one of those where people hear that and think, "but I don't want to displace jobs," or have we passed that now, where people are more comfortable reducing manual steps, automating the process, and improving performance? Nohtal Partansky: I don't think people are so concerned about the job thing. I think the people who don't employ people are concerned about it, but the people who do employ people are past that. They're like, their company's not going to exist if they don't figure out how to reduce the cost of what they're selling, and then no one's going to have a job anyway. So it's really: either you have fewer people working at your company, or you don't have a company. I think people are over that concern. Bryan Fields: I think that's really important to share, because I always feel like when people hear "fully automated," their first thought is, "what about the roles people do?" I think adjusting people into new roles helps improve overall organizational performance. So when you're scoping out that full end-to-end, is it a conversation with the customer about understanding their workflow, and then saying, "here's how we'd do it, here's the entire process," and them saying, "okay, can we automate 80% of this and these two steps stay manual" — or is it more of a back-and-forth throughout the whole workflow? Nohtal Partansky: Those both kind of sound like the same thing, but I'd say it's a little of both. Right now we're working on automating the full line after pre-rolls are manufactured — the full packaging part of the process, not even the pre-roll process itself, just the packaging. We walk through it and say, "okay, we can automate this part, this part, this part," and then stitch those parts together. After that it becomes a matter of diminishing returns, where you could automate a part, but it would cost about as much as all the other parts combined, versus just paying someone $20-30 an hour to do it. Maybe that person is making $50-60k a year doing something simple like loading jars or palletizing, but automating it would cost you $300k — that doesn't make a ton of sense, unless you need three people to do that role because of volume. So it's more of a consulting approach: understanding what's going on, and then getting to the point where the cost-benefit ratio starts tipping against automation — that's usually when you stop. Bryan Fields: What about in cultivation or processing? Are there fully automated partners you're working on there? Nohtal Partansky: Not us. I've seen some facilities that are very automated, but I wouldn't call them automated so much as saying they have a lot of conveyors and people on top of the conveyors. The cultivation side is really hard — I don't know anyone with a fully automated cultivation facility. There was one, in Desert Hot Springs maybe five years ago — I think it was called Sunrise or something like that — a cut-flower botanical greenhouse farm that was retrofitting to grow weed, and they ended up spending $100 million and never got it to work. I do know there are fully automated cut-flower facilities — tulips, lilacs, boutique flowers — where you'll have a million-square-foot facility with maybe 100 people, mostly boxing and handling, but everything moves in this system like a ballet. I haven't seen anything like that for cannabis — I've seen some scaffolding of what that might look like, but nothing fully automated. Bryan Fields: What about robotics in the workforce? Is that something you've been approached about? Nohtal Partansky: Yeah, that's our main thing. There's a big difference between robotics and automation that people don't quite understand. The biggest difference is that a robot will do a set of tasks somewhat independently and then have feedback — the feedback is the most important part. Bryan Fields: Can you give me an example? Nohtal Partansky: Think of a printing press or automated typewriters — they automated steps, but in one direction, with no feedback involved. A digital printer, though, is more like a robot — if it runs into a problem, at least the fancy ones stop. A cheap printer will just print garbage, but a good printer, if it runs out of ink, will tell you; if it jams, it'll tell you; if the lines come out weird, sometimes it has scanners that confirm the print and it'll stop — as opposed to something with none of that feedback. Traditional automation is automation, not robotics — there's no decision-making, very little feedback. An automated facility is like a clockwork mechanism. At my co-packing company, that capper on the jars was automated, not robotic — a jar would come under, a sensor would detect it, and the lid would just spin until it hit the bottom. A robotic system would decide how to start the lid, spin it, place it correctly, and if it's not in the right direction, adjust — continuously modifying its own performance based on feedback from the situation. A lot of things are automated, but not many things are robotic. Bryan Fields: That's a really important distinction, right — the flexibility of a system programmed for one task that needs help when variability comes in, versus robotics that knows how to handle variability and adapt toward a goal. How does the programming work when you're developing this with a customer, to understand the edge cases and make sure it delivers on the production floor? Nohtal Partansky: I don't think customers are involved much in the design of the machines. There's a good Steve Jobs quote — if you asked people in the 1910s what they wanted, they'd say a faster horse and buggy, not a car. You can't really rely on customers to tell you the best way to solve their problem because it's not their domain — they're cannabis manufacturers, not technologists. So it's on the integrator, our company, to deeply understand what they're doing and then build a system that accounts for all the faults and failures. I think that's where a lot of automation companies fail — they don't take into account all that variability, so they don't build it into the design. There's one company that's been building a product for over two years as a competitor to Stardust, and they still don't have something that works, but we built it in six months and have been deploying it for a year and a half — because we took into account a lot of the issues. Even when we deployed our machine the first time, we found all these things we didn't know, but we deploy very early in the design process. We're not trying to polish it until it looks great, because then you overdesign for a small design space when you need to be designing for a large one. That's difficult if you haven't spent a lot of time in a cannabis operation — if you're an engineer visiting a facility for a day or two every couple of months, you don't get a visceral understanding of how different things can be. I spent three years doing it, so it's very different. Bryan Fields: Unpack that a little. When you push a product out at maybe 80% of the way there, and the customer hits the floor with it and provides feedback, do you iterate again, or how does that work? Nohtal Partansky: It's all part of our design process, which we break into a proof-of-concept phase, a prototype phase, a Gen 1 phase, and then a much longer extended phase building toward Gen 2 — which we don't change much and keep for several years. In proof of concept, we try to figure out the worst problems of the application and solve those. Then we build a prototype — something a customer could use, but really a basic model. Then Gen 1 is basically what the final product will be, but it probably won't work — it might break down half the time. We deploy those to one or two customers who we think are competent enough to understand it's still in development and who are friendly with us. That's why we don't make things that are just copies of other companies' products — we want customers okay with it breaking down half the time because it's still five times better than what they had before. Our last products, Chico and Stardust, didn't have competitors when we launched — it took about a year and a half before anyone even built a prototype using us as a foundation. After deploying Gen 1, it takes maybe 6 to 9 months to work out the bugs, including wear-and-tear issues you can't see until then — and it's much easier to solve with five machines deployed than 50. We never just build a bunch and let it rip, because then you spend the next two years fixing things and it's a mess. Bryan Fields: You'd be playing whack-a-mole at that point. Nohtal Partansky: Yeah, yeah. Bryan Fields: So how do you know — let's say you've got it deployed with five customers, getting good feedback from four, but one is giving weird signals. How do you know if that's a symptom of that specific environment versus an actual defect in the product? Nohtal Partansky: If none of the other customers have the issue, there's a pretty high chance it's a customer problem. It's root cause analysis — we're very concerned with truth-seeking: what is actually the problem, where does it come from, and what are the implications? In practice it's hard, because if one customer has things different from everyone else, you might lean toward thinking it's a design problem. You have to have a baseline and compare everything. Nine times out of ten it's the customer doing something differently. We'll go in, ask what they're doing, and in situations where it is a design problem, we discretize every function and input they're using and cross-reference it against the other three or four users. If it's a completely unique situation to that customer, it's probably a symptom of their setup; if not, it's probably a design issue, and we'll fix it — because with a small sample size of five people, you're probably missing things that will come up with more users. In that phase you're not trying to sell, you're trying to design, so you just go, "these guys are having this problem because of their situation, and they probably represent a lot of others who'll eventually get the machine — let's make this change," and see in six months to a year if it comes up again. Usually it does, and then we already have the fix. Bryan Fields: How adaptive are these products to the nuances of different environments? Nohtal Partansky: Stardust is a lot more adaptive. Chico is much more of an automated device than a straight-up robot — I think a lot of the robotic aspect there comes from its connectivity, since the feedback loop isn't onboard, it's through the cloud. Chico is a bit more straightforward, but Stardust can vary quite a bit and change how it behaves. One example: we have a machine learning algorithm for how we heat materials, because we don't know what customers are going to put in the machine or how it will behave — it might be an acacia or glue formulation, distillate, an HTE, or something very terpy, and sometimes they want to heat it and sometimes they don't. All of that behaves very differently when heated, so we built our own machine learning algorithm that approximates the specific heat of whatever is inside and changes how it heats accordingly — kind of like an oven that dynamically adjusts temperature based on how the cookies inside are warming. Bryan Fields: A detail like that probably goes unnoticed by most customers. Nohtal Partansky: Yeah, it's hard to convey the value proposition of something like that. It usually comes down to the reliability of the machines — we're generally first to market with the most market share and the most robust technology, so those little details are one more piece of the puzzle. We'll say, "here's what your end result is going to be," even if we can't explain exactly how each feature contributes — but when they add up, they equal that result. I think we probably should do better at explaining it, but it's almost impossible, because customers will ask, "why can't it just know?" And the answer is, to "just know" means running a machine learning algorithm to understand exactly what's inside and how to handle it, because we don't want to screw up their product by making assumptions. Bryan Fields: It has to be adaptive to exactly what it is. Nohtal Partansky: Yeah, and that's also a problem, because customers see other machines and assume that's how those work too, when really most other machines just heat things up without any machine learning involved. So there are inconsistencies, and it's hard to convey that without having twenty 15-minute conversations explaining exactly how the whole machine works — by the end they'd basically need a bachelor's degree or a Coursera course in mechanical engineering. Bryan Fields: Are there elements in the majority of the cannabis supply chain today that you think could benefit from adopting robotics? Nohtal Partansky: Oh jeez — basically all of it. There's basically no piece of the supply chain that couldn't benefit. Bryan Fields: Let me ask a better question — what's the best way to go about it? Let's say I run a large facility, vertically integrated. I know I have problems across the board — do I start with the most manual, highest-error tasks, or do I start with the lowest-hanging fruit that's easiest to automate? Which approach is best? Nohtal Partansky: That's interesting — I think people approach it from both directions. It really depends on the financial position of the operator. Bryan Fields: Sure, but at the end of the day, say I'm in an okay business position, looking to future-proof and improve, knowing the industry changes fast. Both options are appealing — which gets results faster for my capital? Nohtal Partansky: Then you're comparing two people in the same financial position — what would they do? Most people start from the easiest ones, because those tend to be more traditional and simpler. If you asked me, having run a co-packing company, I would hit the biggest bottlenecks — it requires more money, but it returns more money, meaning it's the highest value. You can see that in how cannabis automation technology has developed since 2020 — there was really nothing before then, and even now there are only a handful of things. Usually the biggest bottlenecks are the first things tackled. On the pre-roll side, the biggest bottleneck was making the pre-roll itself, so I skipped that — I saw a few companies already trying to solve it, and honestly, as of today, I don't think any pre-roll company has fully figured it out. I positioned my company for the next phase — infusion — instead. Similarly with vape cart filling, the first thing attacked was filling — Thompson Duke made one of the first automated fillers — but packaging of vape carts still hasn't really been attacked. From a technology perspective, the highest-value things get attacked first, because no one wants to automate the end of the line while the beginning is still a bottleneck. But sometimes places later in the value chain are actually more costly than the ones tackled first. Take pre-rolls: everyone wanted to automate knockboxing or the pre-roll machine first, but the differential in raw output between a good manual knockbox process and an automated pre-roll machine actually isn't that large. I know people who knockbox 40,000 joints a day with 10 people — to do that with a machine, you'd need eight machines and still need eight people. So there isn't a huge difference, once you maximize the efficiency of your manual process — sometimes there isn't much savings at all, and depending on constraints, you could even go backward. Bryan Fields: That's the root cause analysis — seeking the truth of what's the real bottleneck, and whether the proposed solution actually fixes it or just shifts the problem. Nohtal Partansky: Right, it's about focusing on what your actual bottleneck is now, and what the real solution is. "I'm not making enough pre-rolls" — is the solution to buy a machine, or to improve your current process? Looking at the whole value chain, the differential from semi-automated to manual final packaging is actually much bigger than the differential from knockboxing to a machine. So you'd think, "I should automate this piece first," even though it's toward the end of the line — but that's not how people typically approach it, because they're not strictly evaluating their whole process to see where they could dial things in. Bryan Fields: Going back to basics — I think it's important to talk about the "zero tolerance" environment. What does that mean? Nohtal Partansky: You're talking about the LinkedIn post, right? JPL isn't actually a zero-tolerance environment — it's a low-tolerance environment, like most of aerospace. Cannabis is an agricultural product, so the tolerances are massive by comparison — relative to JPL, you'd probably round the difference to zero. Tolerancing refers to variability from a target. When you machine a part, you never tell a machinist to make a part exactly 2 inches long, because there's so much variability — tool wear, machine vibration — so you specify plus or minus, say, 0.01 inches, which is a fairly large tolerance in manufacturing. In aerospace, you might specify plus or minus 0.0005 inches, five ten-thousandths of an inch, a very tight tolerance. Many people think that's the aerospace standard everywhere, but in cannabis, if you tell someone you need cone length within 0.005 inches, you'll never get it, because cones are made by hand — pull cones from a box of Futurola or Custom Cones and stack 900 of them and they'll look like a wave, not even at all. There's a lot of that in cannabis. It's a different mindset designing for fractions of an inch versus factors of an inch. Bryan Fields: No, that's a good explanation. It demonstrates the experience you had prior, understanding a target needs to be precise, but here there's more flexibility — plus the nuance of people grinding weed differently lot to lot, which all feeds into the pre-roll, and you need to handle that variability. Bryan Fields: Continuing forward — how do you think about protecting IP without slowing down adoption, knowing this industry needs to iterate fast? Nohtal Partansky: We protect IP by filing patents. Before we make anything public, we always file a provisional, which gives us a filing date. You then have one year after filing the provisional to file the non-provisional patent, which generally takes 18 to 24 months to get awarded. Bryan Fields: Is that something you think about often, or is it more "iterate forward and figure out defensibility later"? Nohtal Partansky: It's something we think about pretty often when we make something new. There have been things I've designed that aren't patentable — when I ran my co-packing company I made a brand called Gico Joints and designed a new type of crutch, a half-gram format that was pretty different — I still haven't seen anything else like it on the market. I wanted to patent it, but our IP lawyers said, "dude, this is just a cigarette with a long crutch" — I said, "but no one's doing this," and they said, "I know, but it has to be novel — you just made it bigger." So, okay. We do think about patents both as an offensive weapon and as a defensive shield. There are patent trolls in the cannabis industry too, believe it or not — this industry is a great target for them since it's scrappy with cash and ripe for lawsuits, which is annoying. Stardust has four patents, already publicly filed, and they should be awarded soon; we'll likely file a provisional on our next product before release too. I'm not a very litigious person — I don't like litigation, except when it's unjust. There were knockoff machines, similar to what happened with Makerbot in the 3D printing world — they came out of open source, started making designs, then closed-sourced them, and people still copied them part for part. In situations like that, if someone is using us as their R&D platform, that's unjust, and I'll fight to the ends of the earth. But if someone builds something that accomplishes a similar goal differently — even better, or cheaper — that's just part of the game, and we don't go after them. Bryan Fields: It's a delicate balance for sure. Nohtal Partansky: Yeah, if they're literally copying us checkbox by checkbox, that doesn't make sense to let slide. But if they're just accomplishing the same goal differently, it is what it is. Bryan Fields: What would you build if money or ethics were no obstacle? Nohtal Partansky: Money or ethics no obstacle — what does that mean? Bryan Fields: Like if you could remove any constraint — build a Neuralink for cannabis, whatever — what would you build for the industry if there were no limits? Nohtal Partansky: Okay, well I probably would have made a Neuralink for the cannabis industry — just giving that as an example of putting a computer in something exclusively for this industry. But really, I think I'd try to figure out how to make a good cigarette-style pre-roll machine. I think that's really cool, and it takes a lot of work — I'd estimate low single-digit millions of dollars to get it worked out, and quite a bit of effort given the wide range of variability involved. I want the differential to be large — if someone's doing 40,000 joints a day with 10 people, I'd only want to build something that does 40,000 joints a day with one or two people. That's hard. It's something I've thought about for a long time, but it takes so much time and energy, and I'd only really want to build it for that scale of operator, which is itself a problem, since there aren't many people producing at that volume. I'd probably tackle it maybe five years from now, targeting the consolidated private and public MSOs doing those kinds of numbers. I don't think it's being solved well right now, and it would probably get solved eventually, but it has to make economic sense, and right now it just doesn't. Bryan Fields: Do you think in the next 5 years we'll get full lights-out manufacturing in cannabis facilities? Nohtal Partansky: Definitely not. Bryan Fields: 10 years? Nohtal Partansky: No. Bryan Fields: 20 years? Nohtal Partansky: Yeah, 20 years probably. Bryan Fields: What do you think is the biggest thing holding that up? Nohtal Partansky: No one knows what they're doing, no one has enough money, and no one has enough scale — those are the three things you need to generate any real automated system. If you don't have scale but have a ton of money, you can do whatever you want, but generally scale and money go hand in hand. To do a fully automated lights-out facility right now would cost about $10 million — if you paid my company $10 million over two years, I could build you a fully automated facility end to end, from turkey bags of flower to pallets. But not many people have that money, and those who do often don't really know what they're doing in terms of exact, unchanging SKUs — labeling requirements and rules keep shifting, and you can't build an automated system in a world that's constantly changing, because automation needs a stable foundation to build off of. If the foundation shifts, even a robotic system needs to be reworked. That requires either actual federal rescheduling with clear FDA guidelines, plus another five years of SKU stability, the way a company like Coca-Cola can run the same bottling line because their bottle hasn't changed in ten years, and they also have massive scale, distributing centrally across regions — whereas cannabis is locked state by state, so you'd need $10 million-plus in every single state. Bryan Fields: Right — you'd need scale, and you'd need consolidation, like what Glass House is doing with their greenhouse scale, producing at a fraction of the cost and preparing to distribute if things open up. Nohtal Partansky: Yeah, but they also allegedly sell a lot of their weed into the black market — the production numbers don't really add up when you look at how much flower one facility supposedly produces in a quarter versus what the whole state consumes. The math isn't mathing. My focus is more on the manufacturing side. Cultivation is generally much easier to scale than manufacturing in every situation — it's more about space and a straightforward process, and the variability comes from changing inputs, not from how you handle or package the output. The spread and implications of variability are much larger on the manufacturing side. Cultivation might vary by, say, hours of light — a 50% difference — but on the manufacturing side, a vape cart could go into a bag, a box, get labeled, go into another bag — you've changed the process by 300%. Bryan Fields: Last question — what question do you wish more people asked you? Nohtal Partansky: I thought of a funny one — "can I buy more robots from you?" That'd be a good one. But in terms of my actual customers, I wish more of them asked, "how can I improve my process," whether it pertains to the machines or their company overall, or "how do I train my people better?" — because most of what we sell are high-functioning tools, and if the people using them don't know how to use them properly, things get messed up. "How can I improve what I'm doing" is probably a great question, versus "how do I make this easier" — the end result might be similar, but the mindset behind each is different. Bryan Fields: A hundred percent, I think that's a perfect way to end. So, for our listeners who want to get in touch and learn more about Stardust Robotics, where can they find you? Nohtal Partansky: I'd say LinkedIn is where I'm most active — just search Nohtal, N-T-A-L. And for information about the machines, go to the website, stardustrobotics.com — that's the best place. Bryan Fields: Thanks for taking the time, this was a lot of fun. Nohtal Partansky: Cool, thanks, Bryan.