Love these types of advances though I wonder if weeds and pests will adapt to outwit the robot or if our ability to adapt automation to such changes simply will always/usually be too rapid for nature to keep up.
Pests will adapt to physical pest control like fire or crop rotation by timing their life cycles to avoid it. Robots aren't any different in that regard, if used without care pests could hypothetically adapt to their usage.
The key to preventing this kind of adaptation is multiple modes of action. Robots do offer that ability because they can be used to both physically remove pests, and also chemically treat pests like in the video.
If you have two to three robust modes of action, you can basically prevent resistance from happening entirely.
>It is possible that rye traveled west from Asia Minor as a minor admixture in wheat (possibly as a result of Vavilovian mimicry), and was only later cultivated in its own right.[0]
Also found this about rye while looking that up:
>Recently, scholars have discovered that rye, more than other domesticated crops has followed a weedy species type of domestication process--from wild to weed to crop and then back to weed again.[1]
I learned this from my brother when I was harvesting Job's Tears in the garden. A bunch of foxtail seeds kept landing in the bowl, and looked just about big enough to be a grain - I commented that it seemed likely that some of our grains must have evolved that way, getting selected because they got harder to distinguish from the main plant. He was like "yeah... that's basically the story of Rye" lol. It was something I'd never thought about, but actually trying to harvest a grain made it really clear.
There is a plant that mimics the shapes of other leaves - even if the leaves are made out of plastic. Pretty fascinating, the same photoreceptors suspected to enable this have been found in other plants, but the others don't use it to engage in mimicry.
Weeds adapting to look like non-weeds is highly plausible, IMO.
Boquila trifoliolata can do it within a single generation. Most weeds don't have this capability currently, but a strong selective pressure could make it happen pretty quick with the right lucky mutations.
I'm not saying either of you are wrong but the we'll be needing the robots to do it other ways or 10x as many farmers. It takes a ton of labor to deal with weeds and other issues in a more sustainable way
So I understand this is not realistic currently; However where I currently reside, basically everybody grows their own food, and has plenty of time to be doing other things in the meantime.
So while we all need robots, there are probably a hell of a lot of people who would be able to produce even one crop a year in their backyard, who don't bother.
Not really. Gardening the typical garden crops do - but those are high labor which is why we garden them. You could plant wheat in your garden it would be more labor, but your yields would probably be less, or if you are really good the same.
No reason the robots won't reach "other ways" at a point either. I think it will take a lot of combinations of solutions over time.
I still think there should be a lot more rotation between general crops, animal grazing and replenishment cycles. "Live" soil is becoming a problem and we keep extracting/killing it faster than it's getting replenished. Not to mention all the pesticide issues.
As the video points out (toward the end), since these robots recognize and work on each individual plant, they are much more able to deal with more biodiversity than our current methods, which can only scale for monocrops.
Would be super cool to have an autonomous robot that could get a 'kill list' and destroy every invasive plant in an area while leaving everything else unharmed.
There's "no sign of any weeds/zero biodiversity" and then there's "target the weeds that would crowd out your crop", which is definitely a problem in my raised beds.
I picked up grains from a farm in Dwight, Illinois that makes very limited use of herbicides and pesticides (it seems like oats are all but impossible to grow economically without some use of a weed killing agent). The farmer made the transition for his health and the health of his kids, because he was intuiting that certain health problems were the result of agricultural chemicals. I was struck, walking in their corn field, by the abundance of insects and weeds, and the absence of these quantities in standard ag corn fields struck me as alarming.
On the one hand, the reason for the decline of the various milkweed butterflies appears largely to be a result of areas that once supported some milkweed populations were hit with glyphosate which killed the host plants off at a massive scale. On the other hand this is already happening with chemicals sprayed indiscriminately, this reduces the amount that's sprayed, which could save farmers money, and this reduces glyphosate runoff which is good. Still seems to have the down side that if actually adopted would likely cause even more loss since it could be used on crops where it's avoided now. And no farmer is going to program it to skip milkweed or some other important host plants, they hate milkweed.
Careful usage of cover crops and planting companion plants would be the opposite of that kind of approach. Companion plants will to some degree prevent weeds, generally without competing for the same nutrients as the crop. Usually they're selected based on their ability to attract beneficial insects as well, which function as a form of pest control. This is really only viable at the scale of gardening, but when used effectively, can produce much higher per-area yields.
Having worked a farm for a few summers, my understanding is that we're dealing with the following set of tradeoffs.
1. If you allow weeds, you'll need to till the soil to maintain yields
2. If you till the soil you erode it, damaging the the long term health of your farm and water supplies.
3. If you drop yields globally, people starve (and farmers go out of business)
Since eroding the soil harms biodiversity in a different set of ways, killing weeds seems like the least bad option. Ideally we wouldn't require harsh chemicals to do that but no one has quite figured it out.
No sarcasm, it's true. Most insects are the enemy too.
The philosophy behind modern industrial farming is to basically create a chemical holocaust that extinguishes all plant, animal, and fungal life except the crop.
I'm not surprised you're being downvoted. Us lot in India think the occidental culture is inherently mono-cultural; even with all the DEI stuff, all that is being fostered is phenotypic diversity (the most useless type).
Classical India seems to have had better ways. In an introduction to Dharampal's works (who has written a lot of science in pre-Modern India), Claude Alvares notes how the Dutch, who had reduced the grass species to one or two in their land, had the temerity to advise India, where each backyard has(d) hundreds of species of plant.
Pity, everything in India too is rapidly devolving to Anglo-American monoculture now :( They (the elites, who are the ones that really count) barely even speak their own languages now.
If non-productive plants and insects are mixed in, you will need more farmland to produce the same amount of crops. In the bigger picture it would be counter productive.
That would be great, but I don't think we'll have batteries that can power a tractor for more than a short time on rough and soft terrain anytime soon.
Seems like it should hoe each weed instead of spraying it. I guess that's a little harder to do. Maybe a lot harder.
Two other things I thought were interesting in the video. One, the idea of mapping the whole field is pretty cool. There must be useful things you could do with a millimeter-scale 3D map of an entire field. Two, this is happening in California, just a few hours' drive from Silicon Valley. I hope some of the engineers laid off recently find work in ag tech, because there will be huge opportunities coming with the rise of computer vision and (hopefully soon) robotics.
Small Robot Company have a robot called Dick which uses "Non-Chemical Weeding", i.e. they zap the weed to kill it, and it's root. They use other small robots to do the mapping and planting. I really like the idea of small, specialised robots for different tasks.
Weed zapping is fascinating - but putting the kind of power electronics required to deliver sufficient voltage in a mobile robot is terrifying from an OHSA perspective.
One group I'm familiar with are using 50kv (and "more than enough" amps) for the purpose.
It's not really. Plenty of people have handheld tasers powered by 9v batteries and plenty of farm equipment can (and does) mangle people not being careful around very dangerous machines.
This is just a misunderstanding of the risks involved in various things.
Yeah there are tons of really dangerous things on farms, like Paraquat, or power take-offs. Power doesn't have to be electrical to kill you. Mechanical power will do the job just fine.
I'm well aware that it's far from the only thing that'll kill me on a farm.
That said, comparing a 50Kv shock to a handheld taser (which typically run at 300-500v) makes me think I might not be the one who misunderstands the risks.
You're off by three orders of magnitude. You need about 32 million volts per meter to overcome the resistance of the air. For a gap of a centimeter, it's about a few hundred kilovolts.
And I grew up farming, I'm quite familiar with the dangers.
For example harvesting equipment... the big ones are sort of up to a 30 food wide chainsaw with conveyors and other grabby bits to suck you in. People need to be many meters away from those things while they're running for any sense of safety. A hot spark has nothing on many other dangers and if designed with any sense would only result in severe local burns. Plenty of farm equipment will effortlessly rip a limb off if it doesn't tear your entire body into bite sized chunks.
You know the machines that do that and you don't go near them. When you're the operator you must know where everyone is at all times and know what they mean to do... and be able to communicate with them with a series of gestures, eye contact, short yells, etc. If you don't it's trivially easy to kill your helpers.
Maybe this is a false equivalence, but I think industrial scale leaching of poisons into the water supply is more terrifying and anything that can replace that is worth looking into.
The text "to take place in early 2020" and "AVAILABLE 2021" on that site, in November 2022, does not inspire confidence.
I guess in the near term dozens of specialized robots for each task will be necessary, but the idea I like is teams of general purpose robots using simple tools to perform almost any task "manually". It's too early to implement that now, but I think it will be feasible within my lifetime.
Generality is probably one reason why this robot sprays instead of hoeing. In addition to killing weeds, by changing chemicals this thing can fertilize, and they also mention pollinating and preventing pollination of flowers. (Preventing pollination avoids overburdening some kinds of trees. Currently instead people have to do "thinning" which is simply pulling half of the crop off the tree before it's ripe and throwing it on the ground).
The reason herbicides have to be highly toxic is that they are sprayed indiscriminately, so only small quantities actually get on the target. Farmers don't prefer highly toxic things. They would prefer to use less toxic things, applied directly in large quantities to the target and not elsewhere.
> I saw on one of the illustrations that the spray is made of acetic acid.
The cynic in me wants to point out that this is basically a commerce, so of course they use the best sounding chemical possible. Will they still use that by the time it becomes a commercial product? I guess only time will tell.
Check the SDS for details, you can easily goggle it, though reading them is harder. You put dilute vinegar on your salad. You don't do that for glyphosate because it doesn't add any flavor, but if you put as much on your salad as you do vinegar it would harmless.
Australia's CSIRO has a setup out at Boorowa[0] doing this kind of work. They're doing a lot of traditional research into monitoring and modelling technology, but also doing some neat digital twin style projects.
For tilling to kill weeds, there actually is one that does that, Farmwise[0].
The challenge with Agtech robots in general is that you need to be able to prove payback, including all operations costs, in ideally 12 months or less on equipment.
This is very hard for most crops because you have to be constantly moving from field to field. However, it is doable and I see it happening soon, starting in specific crops (My company L5 [1] is working this for strawberries).
However, as the price of raw tech inputs (robot arms, compute, cameras) continue to fall, I do see lots of “low hanging fruit” that can be won relatively soon, especially with smart software.
I'm glad you mentioned the economics of automating farms. There are numerous products listed in this thread that only appear to target one crop type and only perform one low cost activity (spraying, weeding, etc) that is done once per growing season (and often once per year).
Automation of day-neutral strawberry picking sounded sensible on first glance to me because harvest occurs over many months of a year (extendable by transporting the robots to another region with different growing seasons) thus robots are not sitting around idle for 95% of the year. Additionally, harvesting strawberries is dangerous work for humans to perform due to repetitive twisting, bending and reaching.
However, looking deeper into the economics of strawberries (using Australia as an example), picking costs are stated to only be 20% of fixed growing costs[1] seemingly amounting to AUD$1.30/plant (USD$0.80-1.00/plant). With a typical density of 40000 plants/hectare, robots appear to only be cost neutral if they can be operated at AUD$52000/hectare (USD$30000-40000/hectare) for picking. This results in, for example, two AUD$26000/year (USD$15000-20000/year) robots assigned per hectare needing to pick at a rate of 2 seconds per plant[2] for 12 hours each every day. With 9 million tons of strawberries produced globally per year, the strawberry harvesting market would be AUD$11bn if Australian labour rates, likely amongst the highest, were the global norm.
Would it be fair to say that farm automation will only become viable if everyone accepts the true cost of growing produce is much higher than current prices. That, for example, wages for farm workers would need to be much higher (doubled or more) to be livable wages and also to compensate for injuries and health risks of farm work and an inability to work in manual labour jobs at older age. In Australia we're seeing many news articles on the topic of "lack of farm workers" so a significant upwards movement in farm worker wages seems possible. Or perhaps strict health and safety regulations introduced to protect farm workers from repetitive injuries would result in wide adoption of farm robotics? Or perhaps it will just be cheap robotics that results in mass uptake of farm robotics?.
[2] Per a 2019 example from one of your competitors (https://youtu.be/9cxHYEzMVKQ?t=106) it looks very difficult to pick at a rate of 2 seconds per plant without wastage (e.g. missing obscured berries) but my simple calculations are nonetheless optimistic.
Sydney based startup Agerris was doing mechanical weeding 2 years ago. Unfortunately they went under about a year ago. You can see their platform in action here: https://www.youtube.com/watch?v=TR7FAtgZBzk
> Seems like it should hoe each weed instead of spraying it.
Huh, I actually thought it was using a laser and the 95% chemicals meant through the entire farming process. In that case it's probably much less than 95% chemical use overall, there's still pesticides and fertilizer to consider, not to mention chemicals used in cleaning and packaging.
I wonder what kind of power you'd need to laser weeds at that speed? Now that I think about it, probably like, a lot. And maybe it'd heat up the surrounding plants too much.
Spraying with that much accuracy is still pretty cool though.
Fascinating use of modern computer vision techniques equipped with modern robotics. Challenge of course is scale and getting farms to integrate such features into pre-existing processes over time.
One thought I had while watching that the video didn't seem to explore is that if you can precisely map the field and remember every plant, and you have 4 million plants on your field, then it seems like you could enable precise experiments. Need to find the best mix for fertilizer or pesticides? Seems like you could do very precise and large N experiments to nail down what is best for your plants in your field and possibly eke out some additional return by doing so.
You can go further than that. Monitor every plant's lifecycle and yield, and save the best for seed. How many generations before you get superwheat? You could even have multiple strains for the varying soil conditions across a single field. Add genetic testing, and you can boost genetic diversity for disease resistance. All without attracting 'GMO' labels.
Farmers are already doing that. You don't need to map out each individual plant, as there isn't really much you can do with that level of data. Just map the field down to small squares each covering a few square meters and that is good enough. Going to the individual plant doesn't help as you can't do anything to one plant that won't leak to the next, but even if you could the difference between what two plants next to either other need is insignificant. Across a whole field there is a big difference though, so you want to break the field up some.
They're very smartly having this work as a tractor attachment so they're not reinventing the literal wheels. For small farms, a scaled-down version could be pushed along like a wheelbarrow or attached to something like an ATV or ride-on mower.
But at the home garden scale, I'd like to see a dog-sized spider-like robot that would roam the block doing this. It can be very hard to keep soil weed free, and then weed it once it isn't.
I'd love to work in this industry as a developer... coming from web/cloud what would the average interview be like? Is it more leetcode style? or system design? I would only want intermediate developer position at most for starters. What are some projects that would make someone stand out for AgTech Robotics stuff?
> What are some projects that would make someone stand out for AgTech Robotics stuff?
I guess, in addition to the typical engineering stuff (AI, robotics, etc.) to work with these companies on startup-level, a passion for agriculture, especially historical agriculture, might be what could set you apart from competition.
I interviewed with a few companies in that space last year and they were pretty much standard interviews. Some coding tests, take home projects, the usual stuff. If it looks interesting to you, definitely give it a shot.
Getting in (assuming a typical robotics company/startup) is not that hard, just don't expect to get paid much (relative to most top tier tech compensation). Most of these "fun" fields all either significantly underpay, or are part of subsidiaries like Google X that are next to impossible to get in without exceptional pedigree.
Eg: A chap I work with has tens of thousands of ANOVA grow plots for seed varieties spread across Western Australian Climatic conditions .. he and various assistants routinely fly multispectral camera drones above plot areas and calibration images.
That's an aquisition, store, normalise, process cycle for a fairly steady data stream.
I'm fascinated by Freethink, the company that creates videos like this for tech companies. Has anyone worked with this or similar companies? How much does it cost and how was your experience?
Maybe a dispensing system which more selectively targets the weeds should instead of using less killer, use a less effective killer.
Instead of some clever bit of selective plant killing molecular wizardry, just spray the weed with lye (NaOH). Dissolve the weed and any overspray or residual will lightly raise the pH of the soil. If you have alkaline soil problems, then maybe use sulphuric acid instead.
The agriculture tech industry seems like a really fun and rewarding place to be. I can't wait to see what happens in the future. Would love a job like that.
FarmWise is another company in the space. Their machines offer a targeted cultivation for crops like lettuce, broccoli and cauliflower. See the performance of the system below:
But, am sure having multiple players in the market is a good idea. Yeah, John Deere is obviously a behemoth and may not be as light a nimble as a good startup in advancing this tech! (and hopefully, they wont try to squash out the competition).
Agreed. And was just thinking, to get more farmers to buy these low-usage sprayers, think the government should give incentives to farmers to buy them. Maybe major tax credits for buying them or using less pesticides?
Agree with you that we need to move away from pesticides (and also do think something like lasers sounds like a great idea). And also agree with you that John Deere is not a great example of how to run a company (they are monopolistic and are know to charge insane pricing for repairs).
But, the thing I like is John Deere is the major player in farming equipment. They can have a huge, immediate impact on the amount of pesticide used today - this not a product that is being developed, it is available and already reducing the usage of pesticides.
It'd be better if it came from a better company, but it's a great transition product that will probably have a large impact on pesticide usage.
This is an improvement within the current paradigm of "big ag", but IMHO tech is opening the possibility of a completely different way of doing things. For example, this comment which is below the linked video imagines a robot-maintained food forest, I think it's on point because I, too, believe that the future is swarms of small robots rather than a few big tractors:
> Imagine, a few years down the line, applying these principles to maintaining an edible polyculture ecosystem-farm. Not necessarily a big machine going through a flat field, but small drones tending to trees, vines, herbs, plants, pollinators and water features all in the location that is best for them based on local features. It could micromanage weeding and harvesting, but also composting, planting, grafting and nurturing keystone species and rare species while avoiding pesticide, nitrogen imbalance and soil compaction. A productive farm could look like the garden of Eden.
Oh lovely. OSE is still doing good. I remember reading up on the project a decade or so ago back when it was basically just a wiki and a handful of engineers.
The project is pretty well thought out and extremely ambitious. I can only hope they continue to make good progress on it.
> applying these principles to maintaining an edible polyculture ecosystem-farm
Yup, actually if you read between the edits toward the end of the video I think this is in the minds of those working on it, not sure why it wasn't focused on more, perhaps because it sounds less immediately practical... once the automation is somewhat intelligent, the need for a monoculture for viable large scale agriculture is diminished.
In fact at a certain point of sophistication, especially when harvesting a polyculture becomes viable, it will probably flip and become significantly more profitable long term to abandon monoculture... since the soil health will improve over time which also improves yields.
Also, the key difference between a 'hard' challenge like self-driving cars and something like self-sustaining agriculture is that the reliability doesn't need to be 100%. An achievable target could be something like 'one human to operate the location'. Things will break, software will glitch, maintenance will be required - but that will be OK, as long as the issues can be corrected in a reasonable length of time.
Furthermore, with the right sensors, there's an opportunity for the swarm to self-correct without human intervention based on results. It doesn't have to be exactly right at the moment, because any ill effects will be noticeable and fixable well after a mistake is made.
I feel like what you're describing is still in the realm of sci fi because of limitations in robot autonomy that we have expected to break through many times without success. If we succeed with autonomous driving we will probably succeed in these domains as well. As well as many other domains.
This isn't intuitive to me. If one big tractor has a failure, the whole operation shuts down. If 1/1000 drones breaks, swap in a spare and move it to some kind (potentially automate-able) repair area. If the ideal number of drones is 1000 operationally, you just need to figure out how much of a margin of extra reserves you need to keep things operating. But even with a .1% lower capacity, you'd probably not notice anything, especially compared with 100% lower capacity.
I think the worry is that with 1,000 drones, if you have 10 breaking daily, that's probably a full-time job for someone to be repairing them, especially if they have to go out retrieving them.
One big tractor isn't a full-time job's worth of maintenance.
I guess I don't really understand your assumptions. Why would 1 massive machine capable of doing all the various tasks of managing this farm require less maintenance than 1000 presumably much simpler, smaller machines? Assuming other things besides mobility break, I don't think retrieval would be much more of an issue, or that we'd have reason to expect the failure rate to be higher for each smaller, simpler drone than for the giant, super-complex one. To get the same work done, the behemoth would have to move much faster, have many more specialize tools and sensors all in the same rig, and would not be swappable for regular maintenance without pausing the entire workflow. Small drones could go back to base in shifts (again with some margin of extra drones) so that there are always 1000 actively working.
To be clear: we were talking about a "food forest," not the kind of row-based farm tractor in the video. I certainly agree that the kind of tractor in the video is simpler than the swarm units would be. But a giant tractor that can navigate a forest is a different matter entirely, no?
On large scales, it's more like you have a hundred big tractors and if one of them breaks down the whole operation keeps going because there are 99 big tractors left. The question then becomes: how much maintenance do you have to do per ton of food produced? A swarm of small robots have a lot more parts than a large tractor that does the same amount of work.
There's also energy costs to consider; does a drone swarm use more or less energy than a large tractor for the same result? I'm not sure what the answer is there, and I suppose it depends quite a bit on what kind of work is being done.
I guess it's true that the originally propose 1000 is an arbitrary number, and it would take some experimentation and work to figure out the ideal number of robots per acre to optimize whatever you care about most (yield, power consumption, maintenance costs being obvious contenders). But I think it's easy to say that literally one giant robot covering an entire edible forest is obviously going to be more inefficient and troublesome than a swarm of some size.
and there is the problem of having too much variety on crops and people wanting meat, basically…
farms are pretty automated these days, maybe 100% when tractors can self-repair; which maybe is not _that_ far away… because some already give reports of what went wrong to the central
> But I think it's easy to say that literally one giant robot covering an entire edible forest is obviously going to be more inefficient and troublesome than a swarm of some size.
Yeah, nobody suggested that we have building size tractors so I don't know what you're arguing.
They were talking about "current farming machine sized ones" vs "a drone with some crap attached"
The one big battery is easier and faster - in large part because you can't lift that yourself so you have machines to do the work for you. For AAA you could design the same machine, but odds are you didn't.
If I had to replace 10k batteries, I'd probably make (or buy) a machine for that too. However, even replacing 10k batteries by hand probably costs less in employee hours than a single one-off machine for either example.
(40 hours / 10,000 batteries = 14.4 seconds per battery; USA minimum wage = $7.25/h = $290 per 40h week, about the same as the cheapest hydraulic lift I saw on machinemart.co.uk at current exchange rates, but I'm not a mechanic so I don't know if that would be the right tool for the job anyway).
And of course, the same applies if we really had drone swarms — anything like this would be automated, up to and including full disassembly of damaged units so their parts could be reused for new units, e.g. https://youtu.be/pDZdnbI0MAc
Big tractor isn't more complex than smaller machines. It's just bigger. So you're essentially having 1000x the failure rate.
Big farm will also have more than one machine and the "downtime" will be mostly scheduled maintenance, very little chance of your equipment just deciding to not work today. In case of many smaller farms, well, you can ask the neighbour... from what I remember from living in small village with many farmers "harvest" was often few farms pooling their resources and doing all the fields together, instead of each doing their own.
Taking real life analogy of probably similar complexity scenario. Would you prefer to maintain 1 very big photo copier. Or 100 distributed desktop printers...
'We' have already succeeded in autonomous driving in agricultural and mining domains - these are much simpler, more tightly controlled and less liability prone than open highways and suburban backstreets.
eg:
> We run more than 130 autonomous trucks, part of our Autonomous Haulage System, across our Iron Ore operations. The trucks are operated by a supervisory system and a central controller, rather than a driver. The system uses pre-defined GPS courses to automatically navigate haul roads and intersections and knows actual locations, speeds and directions of all vehicles at all times.
> In 2018, each truck was estimated to have operated on average 700 hours more than conventional haul trucks, with 15% lower costs – delivering clear productivity benefits. They also take truck operators out of harm’s way, reducing the risks associated with working around heavy machinery.
Ehhh really haven't. While you could argue farming is different due to it being more constricted, it really isn't and well unlike driving, this has a size constraint which has and likely will not be solved for in the near or distant future. It's complicated and might not be solvable. And well the current advancements while they seem nice aren't very useful. I know individuals who work in this sphere specifically and there are several problems alone with basic things such as sensors which are not adequate enough, if anyone attempts to solve this problem in the next decade I'd believe them to be the next theranos.
> And well the current advancements while they seem nice aren't very useful.
Current advancements already allow tractors to go autonomously go over the bulk of the same large areas of ground that they have for the past decade (already logged and recorded), graceful fallback alerts remote farmer to make a human decision re: tree falls, new rocks, unexpected gear state changes.
These are useful advancements in machine control.
These developments have locally run in parallel with the above linked Rio Tinto work here in W.Australia - it's a pity the largest mining operation on the globe has imagined its work to be productive and useful and allowed its cost analysis to sway its thinking lacking your insight to the contrary.
My educated guess is detecting humans/children in places where they shouldn't be. Sending a tractor across a field controlling everything is easy - tractors have been doing this for more than 25 years. Detecting something that shouldn't be there and not killing it is hard, which is why all tractors that do this have some form of sensor to ensure someone is in the tractor. There are a few other issues similar to the above.
I work for John Deere, I can't speak for the company.
Working for John Deere I can tell you that we are attempting to solve the problems right now. I can't speak for the company, but that we are working on it is public knowledge.
Did you know that natural swarm dynamics completely baffled us up until a few years ago: How do birds communicate well enough to stay on a group? How do they know which direcion to go? How does it look like a singular consciousness is undulating through hundreds or thousands of individuals? We marvelled and were dumbfounded.
But we had made it so much more complicated because we couldn’t see the few simple rules that each of them were following.
Honestly, I think robot swarms will be simpler than large singles.
> But we had made it so much more complicated because we couldn’t see the few simple rules that each of them were following.
I hope you don't mean "Boids" and the three-four simple rules, because that's just an approximation of a swarm behavior made for artistic purposes - i.e. the result is supposed to pass as a flock of birds or a school of fish to casual onlookers (e.g. gamers, audience of a movie), and not to be accurate to the real thing.
Even if it were "Boids", I don't think it worsens the argument by a lot.
The essence of the statement is that complex behavior can emerge from a simple ruleset.
I'd like to add, that while individual behavior is difficult to predict, a general motion prediction can be much easier. (Think shepherds, for example)
All the complexity in farming is in getting the view on the small fraction of the area and applying treatment (water, fertilizer, removing weeds etc) into that area.
The "swarm" is meaningless there. Each node need to have minimum amount of sensing and computation to handle that job.
It's also at worst "a sloped square" so there is no "swarm mechanic" to move around, just zigzag around the field.
I keep saying I want a laser on a turret that zaps cabbage moths, snails, japanese beetles, squash bugs. Do I need it? No... but it would reduce my labor and losses while increasing profit.
There's so much opportunity for improvement, but we need the R&D money and start-ups to be there too. Please, VCs, spend your money on these projects!
I love this vision and hope it becomes the dominant telos for all of the work underway in this area. In addition to "beautiful farm," imagine a world where farms, parks, wildlife preserves, decarbonization centers, and biological research hubs are all the same place? No reason these swarms can't be optimizing/managing a native ecosystem food-forest with high yield in mind... and likewise easy to imagine unobtrusive, silent robots that don't stand in the way of the ecosystem's normal functioning.
A key problem with small autonomous robots is their energy supply and mass. Batteries have low energy density. Actuators have bad power to weight ratios too. Living things are enormously more efficient. So much so that the intuition apparently offered by watching animals usually does not apply.
This is extremely true for swarm robotics, inspired by ants, termites, etc. Swarm critters rely on the emergent effects of lots of behavior to make up for their limited sensing and computation. By 'limited' I mean relative to a mammal or bird. Lots of behavior means lots of energy spent moving and doing: exactly what is prohibitively expensive for a robot.
Computation and sensing is getting better, smaller and cheaper much faster than batteries and actuators. This trend favors smarter robots that think carefully (cheap) and do the right thing (expensive per shot).
what if we make them bigger than insects but still relatively small? just an arm on a track. heck, we can put the energy in the track too. boom, no battery.
There are some interesting things being done with pivot spans. Hummingbird runs nozzles back and forth on rails suspended from each pivot span. But they’re difficult to keep running.
Ag is hard. It’s dusty. The weather can be brutal. Stuff gets left out year round. The turnover is high. Schedules are dynamic and unforgiving (when the weather turns, you have to adapt your whole operation). There’s not a lot of highly technical staff to maintain things; bailing wire and bandaids is a real thing. The tradition of “hacking things” to keep them working comes from farmers.
The more complicated your machinery, the more your MTBF product goes up. So complicated units don’t scale well.
If a single sprinkler on a pivot fails or misbehaves, there’s nearly 400 of them and there’s some overlap. If your “robot” stalls out with some grit or a power supply issue or a bricked update, or leaky hydraulics, you just lost 1/10 of your pivot. Pray it happens on an inner span, rather than an outer one.
The problem is that swarms of robots can be used for destructive purposes. There, the amount of energy expended and the loss of some drones is no match for the destruction caused. Not to mention all those decentralized self healing swarms are totally resistant to the conventional “destroy the Borg queen” nonsense trope of so many alien movies. (Yes I specifically mentioned the Borg queen because it was the moment Trek nerfed the Borg’s greatest strength and became like all the other denoument alien action flick franchises)
>The problem is that swarms of robots can be used for destructive purposes.
Well, toxic chemical agents can also be used as weapon on human civil, and — sadly — it was actually used that way. Plus even when you use them for farming, you still have pollution that will erode biodiversity, human health, and other sustainable environment negative impacts.
Not to say that robot swarm is necessarily a perfect green solution. You still have to build them and feed them with energy, so even if the robot would be exempt of undesirable side effect on local operational site, the global impact would most likely include plants in their lifecycle.
This could simultaneously absolve the need for autonomous vehicles, because whenever you need to go somewhere, you could just make the giant arm put you there.
That's called a train. It's great. You can read a book while you are transported, it uses very little power to get you there, it can be 0 emission, there are no traffic jams. Sadly it's not fashionable cause you can't show off how wealthy you are by using it.
It's faster than flights on short (<3 hour) routes because you don't have the whole security theatre thing. Also train stations are in the city center usually and airports are in the outskirts.
In the US, most tracks prioritize freight traffic over passenger rail. And there is a tremendous amount of freight traffic. There may not often be traffic jams, but when you are diverted to side track to wait and allow a priority freight train through, it's not uncommon to wait a very long time (I've been delayed more than an hour on the 3 hour train ride from Portland to Seattle).
There are also huge bottlenecks in US track infrastructure that absolutely do lead to traffic jams - perhaps most infamously in New York City. Vetoing the addition of 6 additional tracks through NYC is one of the reason Chris Christie was so hated.
Why not both? You could have each train actually be an array of arms that reach out and pick people up and move them along their route. The arms wouldn't have to be as long since the train is moving along the rail, and a select number of arms could be reserved for first class passengers.
Gigantic arms are not feasible because of how torque and leverage work.
As your arm gets longer, it has to get stronger, thus heavier, thus requiring larger motors, and anchored to the ground better. After a handful of meters, robot arms can’t hold themselves up, let alone carry a useful tool.
There are multiple meter long robot arms for lifting big things in factories but they are enormous, enormously expensive, and consume enormous power.
But here’s a delightful concept that sidesteps this, by making the arm weigh nothing:
We dream of lots of fun things. And even try some of them. But not a few days go by that we don’t once again mutter “power is king” To each other. Everything we do revolves around the constraint of power budgets.
How about having an induction recharge state every few meters? They could be rolled out as the same type of infrastructure as water piping is now. At scale, induction chargers can be manufactured very inexpensively.
Make a square of tensioned cable (or rigid rail) boxing in some area of garden/food forest/farmland at a height above the plants. Have 4 modules which maintain position on opposite sides and holds a tensioned wire between them, making a tensioned wire along the x and a tensioned wire along the y. At the intersection of these wires there is a powered doohicky that can act like a spider and drop down onto the plants below and do its work. The power to the doohicky could be cable from either axis.
It feels like this approach would constrain you from planting in an organic/spontaneous manner; you'd have to plant stuff to make sure it lines up with the cables above right?
Following the other comment re: power being king, what if we used a hub-and-spoke model where small drones with cameras (for sensing) and simple arms w/ clampers (for manipulating) perform 'bursts' of energy intensive activity and aggregate their sensory inputs into one composite report on the health of the plants? They do their thing (maybe for 10 - 15 mins or however long their battery lasts) and then fly back to the hub to charge. The hub itself could rely on a hybrid combo of guaranteed external power and solar panels.
Ofc, it'd be cool to see a cost analysis of how expensive this solution is (from both technical and financial standpoints) versus simpler, more energy-efficient human inputs (aka walk around your garden and prune trees that need it and pick berries that are ripe).
Plus maintenance on large number of things is a b*tch.
My better half owns a small hotel with 19 rooms. Turns out running a simple 10 minutes task in each room takes... entire day (3 hours for the work + switch room overhead + breaks turns out to be full day often).
Not to mention by making it pulled by tractor you get something that complements the existing o̶l̶i̶g̶o̶p̶o̶l̶i̶s̶t̶s̶ large players so it's more like you can do something with them to scale yourself out rather than compete against them to scale yourself out. It's much smarter making something tractor pulled from a business economics perspective, not just from a physics perspective.
Reading the comments in this thread reminded me of "Silent Running" (https://en.wikipedia.org/wiki/Silent_Running). Droids (albeit 70s-style), agro-tech, and post-apocalyptic sci-fi. It's dated but enjoyable.
This sprayer is designed to promote the growth of {everything it doesn't kill}. Those 'weeds' will have as much benefit from the spot-spraying as the crop. Which is either significant, or it's negligible (and this sprayer is useless, so probably not that)
Watching that video made me joyful. I was an agronomist in the early nineties walking soybean fields twelve hours a day mapping weeds and then going back to the office making what's called postemerge chemical recommendations to get most of the weeds that I had observed.
During those hot days having then written software for a hobby I thought there has to be a better way. Could a sprayer map weeds and apply chemicals in a single pass? I did an in depth investigation talking to tech people and university experts to see if what I imagined was possible. What I found out was that we were decades away from the tech making it possible.
A couple of years later Monsanto came out with Roundup resistant soybeans and my soybean field walking was greatly reduced overnight. Problem with this one solution fits all was that soybean yields dipped with the resistant beans for farmers that previously had good weed control and we began to develop more Roundup resistant weeds.
I did notice that this company is working with high value crops only. I believe the reason is the tech currently is quite expensive. Though in time the cost should come down as more farmers use it.
One problem for the Eastern states is the use of shields on the boom. The company that I worked for experimented with shields on our commercial sprayers. In fact we were the first people to do it East of the Rocky Mountains. We bought them because we wanted to be able to spray on windy days. The shields had problems and were parked in the weeds in a single crop season.
The best parts of that job was when you stood still for a minute late in the season, completely oblivious to the stream of fire ants pouring into your boots
I did a lot of soil testing. Refused to go out on November 15th, the start of deer season, after the 4WD pickup I was using was shot up in a field. I hugged the floorboard as shots went through the truck and just over me. How a bright yellow pickup could be confused with a deer is beyond me.
That would be a good time to use those extra-loud non-stop alarm in your car.
Every single people working in their property should have one working non-stop in the deer season to signal that is not a deer. Some just are deertonic and unable to distinguish it from cars.
>How a bright yellow pickup could be confused with a deer is beyond me.
It wasn't. It was confused with a government vehicle. The one day a year that is synonymous with hunting accidents just provided plausible deniability for taking a few pot shots.
We had such a system. It was called slavery. Lots of cheap labor. And still we arranged our crops in the style of vineyards, so as to reduce to travel time of workers between plants.
Yes, the current ways of doing intensive farming are hard on the land and ultimately very wasteful. If you look at the progress in pesticide free organic farming and indoor/vertical farming, what they have in common is that they leverage a more data driven approach where you simply don't need to pre-emptively correct problems if you can manage the environment in which you grow plants better. Avoiding the conditions that cause pests and weeds to get out control is much more efficient than dealing with them after things get out of control.
Basically, intensive agriculture is compensating for its destructive nature by just brutally suppressing any "undesirable" growth. In nature, pests target the weaker plants. Basically, you stress out plants by mismanaging them and they are fair game for a wide range of fungi, insects, etc. E.g. under or over watering them, depleting the soil, etc. all stress plants out. If you have unhealthy soil, it's hard to have healthy plants growing in it.
Intensive agriculture starts with disrupting the soil (by plowing it). This causes the soil to degrade over time and releases captured carbon into the atmosphere. The natural response is for aggressive weeds to start growing in disrupted soil. To "fix" this, farmers use a mix of pesticides and fertilizers. The pesticides kill a lot more than just the pests and weeds and then further degrades the soil and the bio diversity in and on it. And finally the fertilizer is just bleeding out of the land straight into the ground water and surrounding areas. It works but it is resource intensive. You need lots of water, chemicals, energy, etc to grow a single crop at the cost of literally anything else. Basically modern intensive farming grows mono cultures on top of artificially created deserts.
Organic farming avoids doing that. Try not to disrupt the soil. Kill off (and compost) weaker plants rather than trying to rescue them. Use nature to keep pests and weeds in check. Etc. It works but it is more labor intensive. You need people to look after things. You get healthier and tastier produce (cooks love organically grown stuff for this reason). However, automating that labor makes it more scalable. That's the opportunity here.
> If you look at the progress in pesticide free organic farming and indoor/vertical farming
By what measure is vertical farming more efficient?.
I've heard the contrary regarding green house gas emissions, it is still less efficient in most crops, it consumes too much power and that overcompensates any efficiency gained in other areas
It's interesting that this conclusion can change drastically if electricity becomes 10-100x more efficient
I don't know what you've heard and where you've heard it but there is generally a lot of misinformation floating around and most of it is being sponsored by lobbies with an agenda to keep the oil and gas companies going. So, I would suggest not taking any of that at face value. Including what you hear in threads like this or from me.
Vertical farms are essentially closed systems where nutrients, energy and some water goes in and produce comes out. There's a bit more to it of course but that's the general idea. They are climate controlled environments where light is provided via grow lights. In short, it uses energy which needs to come from somewhere.
You are confusing power requirements and greenhouse gases. Mostly these things use some form of sustainable power. Either sourced indirectly via some electricity provider; or directly with e.g. solar panels. If you think about it, the number one expensive consumable here is electricity and the cheapest way to produce it is via renewables. It's that simple. Why would you get more expensive energy if you can get it cheaper. Keeping all those ACs, lights, pumps, etc. going you are going to have to figure that one out or deal with competition that is smarter than you on that front.
There are probably a few farms that are forced to buy via the grid and are hopelessly dependent on their local coal burning monopolists. Not a great business plan but the good news is that coal is rapidly becoming a very unfashionable way to produce power due to the fact that it is expensive and dirty. It being expensive cuts down on energy margins and causes coal dependent producers to be dealing with awkward things like going bankrupt and being forced to pay more for their coal than they can charge for the resulting energy when wind is blowing and the sun is shining. Gas has similar issues, as people that pay gas bills these days can vouch for. Mine went up by 2x recently. Not great.
> They are climate controlled environments where light is provided via grow lights
Solar's efficiency is below 50% [https://www.nrel.gov/pv/cell-efficiency.html]. Considering that, I'd guess for each m² of solar panel deployed, you'll get <50% of the yield that would be produced in that same land with horizontal farming (assuming fertile land).
You can argue that solar panels can be deployed in non-agricultural surfaces, and that's true, that allows to grow things where it was not possible before. But I'm not convinced that vertical farming is more efficient than horizontal on very important metrics like cost and GHG emissions
It doesn't even need to use herbicides because it removes weeds mechanically. We actually build it to also fit into Europe's small(er) scale farms as well.
Regarding the energy issues mentioned multiple times in various threads here: We managed to make it solar powered and battery buffered to work autonomously for more than a working day.
If only there were swarms of self replicating autonomous systems that can heal, recharge and defend themselves without input and adapt to new situations, while being stress tested for free for millions of years.
But I'm dreaming, certainly if it would have existed humans would have tried to leverage them a long time ago.
Snark aside, as useful as the birds and the bees are, they've got their own goals, and those goals aren't always aligned with ours — hence the use of pesticides and insecticides, or indeed scarecrows, in the first place.
> the future is swarms of small robots rather than a few big tractors:
You are missing something important: soil compaction is a sub linear function of weight. which means that larger tractors that cover more area with the same tire tracks are must friendly for the soil, compared to many small tractors. Because of this factor larger is better. Eventually we may get to the point where the tractor runs on tracks, to allow even bigger tractors that touch a smaller are of soil.
Sure flying drones will have less soil compaction, and they will have a place in the future, but keeping something in the air uses vastly more energy than wheels on the ground.
I feel this dream of robots everywhere gets us farther from a garden of Eden.
To cultivate the earth we're going to mine other parts of it? I'm not suggesting getting rid of the tools that improve quality of life but does automating one aspect not cause 100 other things down the line? There's mining, factories, repairing, replacing.
Maybe instead of robots we work on getting people to have gardens on their property, in cities use the rooftops, stop importing foods from across the globe, etc.
When I see stuff like this sometimes I get depressed that I'm building apps that just move little bits of information around the web and don't have any real demonstrable value to the world.
How hard is the jump into the realm of computer vision/AI/robotics for someone with strong skills and experience building web applications? Is the coding much different? Is a company like this employing math/theroy-heavy people to develop new types of AI or are they working higher up the stack and just need people who can pipe data around?
Positions do all of the above exist. Just depends on the company and where in the stack you are. There's a lot of math and domain expertise on the controls and sensors sides. Most people in any particular company aren't doing those though. There are lots of jobs available for c/c++/Matlab/Python/js.
If you don't want to change what you're doing or using, most modern robotic platforms have a web interface somewhere in them for monitoring the fleet. Any web developer would find their skills perfectly compatible.
It's a common misconception that you need a ML or robotics PhD to work in the field. There are plenty of frontend & backend web engineering, data engineering, and infrastructure/devops roles available at robotics companies - you don't need to be a domain expert.
For example, at Foxglove[0] we are building open source web-based visualization and data management for robotics (shameless plug: currently hiring).
> How hard is the jump into the realm of computer vision/AI/robotics for someone with strong skills and experience building web applications?
If you are willing to make the jump. It's possible. You have an advantage.
> Is the coding much different?
Not as much as they have marketed you to believe.
> Is a company like this employing math/theroy-heavy people to develop new types of AI or are they working higher up the stack and just need people who can pipe data around?
No. They are not making academic thesis. You'd be surprised how primitive and "basic" many of these companies are (comparing to their marketing videos, which are click/eye baits). You can definitively make a difference there.
Of course, as other comments have mentioned, these companies still need front-end, databases, backend, and you know, normal basic automation scripts like everyone else.
Robotics is a big field and encompasses a huge range of things. Similarly a lot of computer vision is pretty technical, particularly for robotics applications. Have a look at ROS and play around, you can simulate a lot of stuff. I'd also suggest Nvidia's Jetbot but the chip shortage has killed it.
AI-wise, nowadays model training stacks are practically codeless. Training a state of the art model on custom data usually boils down to changing some parameters in a config file, plus some boilerplate code. Not always that easy of course, but training an image classifier today can be a single line of Python.
It's not so much piping data around but dealing with datasets in any kind of non-trivial problem is where the hard work is. Neural nets are just another tool for most data scientists. For example exploring Kaggle will get you familiar with weird datasets, but actually collecting, arranging for data to be labelled, and then cleaning it is a skill on its own.
That's usually where experience comes in. Hence there are a lot of startups charging frankly obscene amounts of money for data labelling and curation tools.
I work at a company doing smart stuff with computers (which some would consider AI) that helps with, among others, the energy transition and getting people across the world broadband. Only about 2 out of 20 devs are actually working on the complicated algorithms. The others are working on the stuff you need to build around it to allow people to actually use the algos. On top of the algos is a pretty complicated web app, which takes most of the actual development time.
So you won't actually have to jump into a whole new career in order to contribute to these types of products. You just have to search for a company that works on a challenge you want to contribute to.
It's very much possible, the range of skills needed to deploy actual ML-driven applications ranges from web developers, data engineers, ML engineers, Ops guys to data scientists, so pick what you like.
Then again: Speaking from experience, most uses cases for ML are not fancy, world-saving new approaches like the one in the video but revolve around quite mundane applications and tasks like "how to make users click on this add", "how much will we sell next month", "can we automate the creation of thix Excel-file" etc., so pick your company carefully
Creator of Django did an ML project last year and has a lot of good insights - general web / programming know-how is really useful for building custom tooling to manage / observe datasets and things like that.
325 comments
[ 3.8 ms ] story [ 299 ms ] threadThe key to preventing this kind of adaptation is multiple modes of action. Robots do offer that ability because they can be used to both physically remove pests, and also chemically treat pests like in the video.
If you have two to three robust modes of action, you can basically prevent resistance from happening entirely.
>It is possible that rye traveled west from Asia Minor as a minor admixture in wheat (possibly as a result of Vavilovian mimicry), and was only later cultivated in its own right.[0]
Also found this about rye while looking that up:
>Recently, scholars have discovered that rye, more than other domesticated crops has followed a weedy species type of domestication process--from wild to weed to crop and then back to weed again.[1]
[0]https://en.wikipedia.org/wiki/Rye
[1]https://www.thoughtco.com/rye-the-domestication-history-4092...
- going invisible?
Weeds adapting to look like non-weeds is highly plausible, IMO.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8903786/#:~:tex....
It's not highly plausible. It would take a lot of generations and a lot of mutations for this to happen.
But really, the whole idea of having perfectly manicured fields like that with no sign of any weeds around seems like the real problem to me.
Zero biodiversity…we wonder where the insects are going ?
So while we all need robots, there are probably a hell of a lot of people who would be able to produce even one crop a year in their backyard, who don't bother.
I still think there should be a lot more rotation between general crops, animal grazing and replenishment cycles. "Live" soil is becoming a problem and we keep extracting/killing it faster than it's getting replenished. Not to mention all the pesticide issues.
Could be a fantastic diversity-enabling tool.
1. If you allow weeds, you'll need to till the soil to maintain yields
2. If you till the soil you erode it, damaging the the long term health of your farm and water supplies.
3. If you drop yields globally, people starve (and farmers go out of business)
Since eroding the soil harms biodiversity in a different set of ways, killing weeds seems like the least bad option. Ideally we wouldn't require harsh chemicals to do that but no one has quite figured it out.
Yes.
>Zero biodiversity
Good.
The philosophy behind modern industrial farming is to basically create a chemical holocaust that extinguishes all plant, animal, and fungal life except the crop.
Classical India seems to have had better ways. In an introduction to Dharampal's works (who has written a lot of science in pre-Modern India), Claude Alvares notes how the Dutch, who had reduced the grass species to one or two in their land, had the temerity to advise India, where each backyard has(d) hundreds of species of plant.
https://www.youtube.com/watch?v=B27IBNotwpE
Pity, everything in India too is rapidly devolving to Anglo-American monoculture now :( They (the elites, who are the ones that really count) barely even speak their own languages now.
Two other things I thought were interesting in the video. One, the idea of mapping the whole field is pretty cool. There must be useful things you could do with a millimeter-scale 3D map of an entire field. Two, this is happening in California, just a few hours' drive from Silicon Valley. I hope some of the engineers laid off recently find work in ag tech, because there will be huge opportunities coming with the rise of computer vision and (hopefully soon) robotics.
https://www.smallrobotcompany.com/weed-killing
One group I'm familiar with are using 50kv (and "more than enough" amps) for the purpose.
If people have to stay around to care for it, then yes, it's scary.
This is just a misunderstanding of the risks involved in various things.
That said, comparing a 50Kv shock to a handheld taser (which typically run at 300-500v) makes me think I might not be the one who misunderstands the risks.
You're off by three orders of magnitude. You need about 32 million volts per meter to overcome the resistance of the air. For a gap of a centimeter, it's about a few hundred kilovolts.
And I grew up farming, I'm quite familiar with the dangers.
For example harvesting equipment... the big ones are sort of up to a 30 food wide chainsaw with conveyors and other grabby bits to suck you in. People need to be many meters away from those things while they're running for any sense of safety. A hot spark has nothing on many other dangers and if designed with any sense would only result in severe local burns. Plenty of farm equipment will effortlessly rip a limb off if it doesn't tear your entire body into bite sized chunks.
You know the machines that do that and you don't go near them. When you're the operator you must know where everyone is at all times and know what they mean to do... and be able to communicate with them with a series of gestures, eye contact, short yells, etc. If you don't it's trivially easy to kill your helpers.
in my experience, this is the best approach. I worked at a few biotechs that did a lot of lab automation.
the robots that could do everything were hardly ever used because they did everything poorly and seemed to be constantly in need of repair.
the robots that could do one simple job were always in use.
I guess in the near term dozens of specialized robots for each task will be necessary, but the idea I like is teams of general purpose robots using simple tools to perform almost any task "manually". It's too early to implement that now, but I think it will be feasible within my lifetime.
Generality is probably one reason why this robot sprays instead of hoeing. In addition to killing weeds, by changing chemicals this thing can fertilize, and they also mention pollinating and preventing pollination of flowers. (Preventing pollination avoids overburdening some kinds of trees. Currently instead people have to do "thinning" which is simply pulling half of the crop off the tree before it's ripe and throwing it on the ground).
Well targeted as they seem to do, I guess it is a good product to preserve the soil and wanted plants with no harmful residues.
The cynic in me wants to point out that this is basically a commerce, so of course they use the best sounding chemical possible. Will they still use that by the time it becomes a commercial product? I guess only time will tell.
I wonder about the effects of acidification after long term use. Perhaps some alkaline filler can be added to balance.
I put vinegar in my salad sauce, I would not dare with glyphosate.
[0]: https://www.csiro.au/en/work-with-us/industries/agriculture/...
The challenge with Agtech robots in general is that you need to be able to prove payback, including all operations costs, in ideally 12 months or less on equipment.
This is very hard for most crops because you have to be constantly moving from field to field. However, it is doable and I see it happening soon, starting in specific crops (My company L5 [1] is working this for strawberries).
However, as the price of raw tech inputs (robot arms, compute, cameras) continue to fall, I do see lots of “low hanging fruit” that can be won relatively soon, especially with smart software.
[0] https://farmwise.io/ [1] https://www.l5automation.com/
Automation of day-neutral strawberry picking sounded sensible on first glance to me because harvest occurs over many months of a year (extendable by transporting the robots to another region with different growing seasons) thus robots are not sitting around idle for 95% of the year. Additionally, harvesting strawberries is dangerous work for humans to perform due to repetitive twisting, bending and reaching.
However, looking deeper into the economics of strawberries (using Australia as an example), picking costs are stated to only be 20% of fixed growing costs[1] seemingly amounting to AUD$1.30/plant (USD$0.80-1.00/plant). With a typical density of 40000 plants/hectare, robots appear to only be cost neutral if they can be operated at AUD$52000/hectare (USD$30000-40000/hectare) for picking. This results in, for example, two AUD$26000/year (USD$15000-20000/year) robots assigned per hectare needing to pick at a rate of 2 seconds per plant[2] for 12 hours each every day. With 9 million tons of strawberries produced globally per year, the strawberry harvesting market would be AUD$11bn if Australian labour rates, likely amongst the highest, were the global norm.
Would it be fair to say that farm automation will only become viable if everyone accepts the true cost of growing produce is much higher than current prices. That, for example, wages for farm workers would need to be much higher (doubled or more) to be livable wages and also to compensate for injuries and health risks of farm work and an inability to work in manual labour jobs at older age. In Australia we're seeing many news articles on the topic of "lack of farm workers" so a significant upwards movement in farm worker wages seems possible. Or perhaps strict health and safety regulations introduced to protect farm workers from repetitive injuries would result in wide adoption of farm robotics? Or perhaps it will just be cheap robotics that results in mass uptake of farm robotics?.
[1] https://www.abc.net.au/news/rural/2022-06-25/breakdown-of-th...
[2] Per a 2019 example from one of your competitors (https://youtu.be/9cxHYEzMVKQ?t=106) it looks very difficult to pick at a rate of 2 seconds per plant without wastage (e.g. missing obscured berries) but my simple calculations are nonetheless optimistic.
Huh, I actually thought it was using a laser and the 95% chemicals meant through the entire farming process. In that case it's probably much less than 95% chemical use overall, there's still pesticides and fertilizer to consider, not to mention chemicals used in cleaning and packaging.
I wonder what kind of power you'd need to laser weeds at that speed? Now that I think about it, probably like, a lot. And maybe it'd heat up the surrounding plants too much.
Spraying with that much accuracy is still pretty cool though.
The attachment they built needed a custom cooling rig to deal with the dozens of lasers. Super interesting engineering going on with agtech.
https://www.freethink.com/robots-ai/farming-robot
Seems like a phased array microwave would be cheaper, require no moving parts, and be able to target the root system ... and be less of a safety risk.
To get the spraying done by firing up a few drones would be excellent.
Particularly with weeds that are in hard to reach places on the farm.
And, you know, cheap. And usable in difficult terrain.
I can't imagine my father, or any farmer who does their job at small scales here in developing countries, being able to afford technology like that.
Yup. I've got four acres of grass that I'd love to farm... without having the farm be a full time job.
But at the home garden scale, I'd like to see a dog-sized spider-like robot that would roam the block doing this. It can be very hard to keep soil weed free, and then weed it once it isn't.
I guess, in addition to the typical engineering stuff (AI, robotics, etc.) to work with these companies on startup-level, a passion for agriculture, especially historical agriculture, might be what could set you apart from competition.
Eg: A chap I work with has tens of thousands of ANOVA grow plots for seed varieties spread across Western Australian Climatic conditions .. he and various assistants routinely fly multispectral camera drones above plot areas and calibration images.
That's an aquisition, store, normalise, process cycle for a fairly steady data stream.
https://www.freethink.com/about/for-brands
Another interesting Freethink video: "Generalist vs. specialist: Which Is Better?". https://www.youtube.com/watch?v=ER2R-F68L0c
Instead of some clever bit of selective plant killing molecular wizardry, just spray the weed with lye (NaOH). Dissolve the weed and any overspray or residual will lightly raise the pH of the soil. If you have alkaline soil problems, then maybe use sulphuric acid instead.
https://youtu.be/vtM_oSiuxeM
But, am sure having multiple players in the market is a good idea. Yeah, John Deere is obviously a behemoth and may not be as light a nimble as a good startup in advancing this tech! (and hopefully, they wont try to squash out the competition).
> pesticide sprayer
Says all about the company.
Lasers are a MUCH better option. John Deere is not a reputable company when it comes to farming.
But, the thing I like is John Deere is the major player in farming equipment. They can have a huge, immediate impact on the amount of pesticide used today - this not a product that is being developed, it is available and already reducing the usage of pesticides.
It'd be better if it came from a better company, but it's a great transition product that will probably have a large impact on pesticide usage.
> Imagine, a few years down the line, applying these principles to maintaining an edible polyculture ecosystem-farm. Not necessarily a big machine going through a flat field, but small drones tending to trees, vines, herbs, plants, pollinators and water features all in the location that is best for them based on local features. It could micromanage weeding and harvesting, but also composting, planting, grafting and nurturing keystone species and rare species while avoiding pesticide, nitrogen imbalance and soil compaction. A productive farm could look like the garden of Eden.
The project is pretty well thought out and extremely ambitious. I can only hope they continue to make good progress on it.
Yup, actually if you read between the edits toward the end of the video I think this is in the minds of those working on it, not sure why it wasn't focused on more, perhaps because it sounds less immediately practical... once the automation is somewhat intelligent, the need for a monoculture for viable large scale agriculture is diminished.
In fact at a certain point of sophistication, especially when harvesting a polyculture becomes viable, it will probably flip and become significantly more profitable long term to abandon monoculture... since the soil health will improve over time which also improves yields.
Also, the key difference between a 'hard' challenge like self-driving cars and something like self-sustaining agriculture is that the reliability doesn't need to be 100%. An achievable target could be something like 'one human to operate the location'. Things will break, software will glitch, maintenance will be required - but that will be OK, as long as the issues can be corrected in a reasonable length of time.
One big tractor isn't a full-time job's worth of maintenance.
Now, the video is shot in a shiny nice weather day. Now imagine how would it perform if it's raining since last Tuesday.
There's also energy costs to consider; does a drone swarm use more or less energy than a large tractor for the same result? I'm not sure what the answer is there, and I suppose it depends quite a bit on what kind of work is being done.
farms are pretty automated these days, maybe 100% when tractors can self-repair; which maybe is not _that_ far away… because some already give reports of what went wrong to the central
Yeah, nobody suggested that we have building size tractors so I don't know what you're arguing.
They were talking about "current farming machine sized ones" vs "a drone with some crap attached"
The only reason you think it's easy is because you ignore the number and just imagine doing it once anyway.
(40 hours / 10,000 batteries = 14.4 seconds per battery; USA minimum wage = $7.25/h = $290 per 40h week, about the same as the cheapest hydraulic lift I saw on machinemart.co.uk at current exchange rates, but I'm not a mechanic so I don't know if that would be the right tool for the job anyway).
And of course, the same applies if we really had drone swarms — anything like this would be automated, up to and including full disassembly of damaged units so their parts could be reused for new units, e.g. https://youtu.be/pDZdnbI0MAc
Big farm will also have more than one machine and the "downtime" will be mostly scheduled maintenance, very little chance of your equipment just deciding to not work today. In case of many smaller farms, well, you can ask the neighbour... from what I remember from living in small village with many farmers "harvest" was often few farms pooling their resources and doing all the fields together, instead of each doing their own.
'We' have already succeeded in autonomous driving in agricultural and mining domains - these are much simpler, more tightly controlled and less liability prone than open highways and suburban backstreets.
eg:
> We run more than 130 autonomous trucks, part of our Autonomous Haulage System, across our Iron Ore operations. The trucks are operated by a supervisory system and a central controller, rather than a driver. The system uses pre-defined GPS courses to automatically navigate haul roads and intersections and knows actual locations, speeds and directions of all vehicles at all times.
> In 2018, each truck was estimated to have operated on average 700 hours more than conventional haul trucks, with 15% lower costs – delivering clear productivity benefits. They also take truck operators out of harm’s way, reducing the risks associated with working around heavy machinery.
- https://www.riotinto.com/en/about/innovation/automation
Current advancements already allow tractors to go autonomously go over the bulk of the same large areas of ground that they have for the past decade (already logged and recorded), graceful fallback alerts remote farmer to make a human decision re: tree falls, new rocks, unexpected gear state changes.
These are useful advancements in machine control.
These developments have locally run in parallel with the above linked Rio Tinto work here in W.Australia - it's a pity the largest mining operation on the globe has imagined its work to be productive and useful and allowed its cost analysis to sway its thinking lacking your insight to the contrary.
I work for John Deere, I can't speak for the company.
But we had made it so much more complicated because we couldn’t see the few simple rules that each of them were following.
Honestly, I think robot swarms will be simpler than large singles.
I hope you don't mean "Boids" and the three-four simple rules, because that's just an approximation of a swarm behavior made for artistic purposes - i.e. the result is supposed to pass as a flock of birds or a school of fish to casual onlookers (e.g. gamers, audience of a movie), and not to be accurate to the real thing.
I'd like to add, that while individual behavior is difficult to predict, a general motion prediction can be much easier. (Think shepherds, for example)
The "swarm" is meaningless there. Each node need to have minimum amount of sensing and computation to handle that job.
It's also at worst "a sloped square" so there is no "swarm mechanic" to move around, just zigzag around the field.
There's so much opportunity for improvement, but we need the R&D money and start-ups to be there too. Please, VCs, spend your money on these projects!
A key problem with small autonomous robots is their energy supply and mass. Batteries have low energy density. Actuators have bad power to weight ratios too. Living things are enormously more efficient. So much so that the intuition apparently offered by watching animals usually does not apply.
This is extremely true for swarm robotics, inspired by ants, termites, etc. Swarm critters rely on the emergent effects of lots of behavior to make up for their limited sensing and computation. By 'limited' I mean relative to a mammal or bird. Lots of behavior means lots of energy spent moving and doing: exactly what is prohibitively expensive for a robot.
Computation and sensing is getting better, smaller and cheaper much faster than batteries and actuators. This trend favors smarter robots that think carefully (cheap) and do the right thing (expensive per shot).
Ag is hard. It’s dusty. The weather can be brutal. Stuff gets left out year round. The turnover is high. Schedules are dynamic and unforgiving (when the weather turns, you have to adapt your whole operation). There’s not a lot of highly technical staff to maintain things; bailing wire and bandaids is a real thing. The tradition of “hacking things” to keep them working comes from farmers.
The more complicated your machinery, the more your MTBF product goes up. So complicated units don’t scale well.
If a single sprinkler on a pivot fails or misbehaves, there’s nearly 400 of them and there’s some overlap. If your “robot” stalls out with some grit or a power supply issue or a bricked update, or leaky hydraulics, you just lost 1/10 of your pivot. Pray it happens on an inner span, rather than an outer one.
It’s interesting to note how the size of Meganeura is credited to environmental factors (oxygen ration in the air). Context is everything, as usual.
https://en.wikipedia.org/wiki/Meganeura
Well, toxic chemical agents can also be used as weapon on human civil, and — sadly — it was actually used that way. Plus even when you use them for farming, you still have pollution that will erode biodiversity, human health, and other sustainable environment negative impacts.
Not to say that robot swarm is necessarily a perfect green solution. You still have to build them and feed them with energy, so even if the robot would be exempt of undesirable side effect on local operational site, the global impact would most likely include plants in their lifecycle.
In Europe, taking the train is often more expensive than flying.
Sadly it's not fashionable because it isn't cheaper and takes a lot longer.
In the US, most tracks prioritize freight traffic over passenger rail. And there is a tremendous amount of freight traffic. There may not often be traffic jams, but when you are diverted to side track to wait and allow a priority freight train through, it's not uncommon to wait a very long time (I've been delayed more than an hour on the 3 hour train ride from Portland to Seattle).
There are also huge bottlenecks in US track infrastructure that absolutely do lead to traffic jams - perhaps most infamously in New York City. Vetoing the addition of 6 additional tracks through NYC is one of the reason Chris Christie was so hated.
https://www.amtrak.com/privately-owned-rail-cars
There was even some hn post on it where someone detailed how much it costs and what you get
As your arm gets longer, it has to get stronger, thus heavier, thus requiring larger motors, and anchored to the ground better. After a handful of meters, robot arms can’t hold themselves up, let alone carry a useful tool.
There are multiple meter long robot arms for lifting big things in factories but they are enormous, enormously expensive, and consume enormous power.
But here’s a delightful concept that sidesteps this, by making the arm weigh nothing:
https://youtu.be/INTHRNcyW9w
I work in this industry (contributing architect of https://nelsonirrigation.com/products/twig-wireless-controls...).
We dream of lots of fun things. And even try some of them. But not a few days go by that we don’t once again mutter “power is king” To each other. Everything we do revolves around the constraint of power budgets.
Make a square of tensioned cable (or rigid rail) boxing in some area of garden/food forest/farmland at a height above the plants. Have 4 modules which maintain position on opposite sides and holds a tensioned wire between them, making a tensioned wire along the x and a tensioned wire along the y. At the intersection of these wires there is a powered doohicky that can act like a spider and drop down onto the plants below and do its work. The power to the doohicky could be cable from either axis.
Following the other comment re: power being king, what if we used a hub-and-spoke model where small drones with cameras (for sensing) and simple arms w/ clampers (for manipulating) perform 'bursts' of energy intensive activity and aggregate their sensory inputs into one composite report on the health of the plants? They do their thing (maybe for 10 - 15 mins or however long their battery lasts) and then fly back to the hub to charge. The hub itself could rely on a hybrid combo of guaranteed external power and solar panels.
Ofc, it'd be cool to see a cost analysis of how expensive this solution is (from both technical and financial standpoints) versus simpler, more energy-efficient human inputs (aka walk around your garden and prune trees that need it and pick berries that are ripe).
This technology has existed for a long time now (see cable robots, or cable-suspended cameras)
My better half owns a small hotel with 19 rooms. Turns out running a simple 10 minutes task in each room takes... entire day (3 hours for the work + switch room overhead + breaks turns out to be full day often).
Of course, that movie does have certain other implications and side effects....
And very, very sad.
This sprayer is designed to promote the growth of {everything it doesn't kill}. Those 'weeds' will have as much benefit from the spot-spraying as the crop. Which is either significant, or it's negligible (and this sprayer is useless, so probably not that)
During those hot days having then written software for a hobby I thought there has to be a better way. Could a sprayer map weeds and apply chemicals in a single pass? I did an in depth investigation talking to tech people and university experts to see if what I imagined was possible. What I found out was that we were decades away from the tech making it possible.
A couple of years later Monsanto came out with Roundup resistant soybeans and my soybean field walking was greatly reduced overnight. Problem with this one solution fits all was that soybean yields dipped with the resistant beans for farmers that previously had good weed control and we began to develop more Roundup resistant weeds.
I did notice that this company is working with high value crops only. I believe the reason is the tech currently is quite expensive. Though in time the cost should come down as more farmers use it.
One problem for the Eastern states is the use of shields on the boom. The company that I worked for experimented with shields on our commercial sprayers. In fact we were the first people to do it East of the Rocky Mountains. We bought them because we wanted to be able to spray on windy days. The shields had problems and were parked in the weeds in a single crop season.
I did a lot of soil testing. Refused to go out on November 15th, the start of deer season, after the 4WD pickup I was using was shot up in a field. I hugged the floorboard as shots went through the truck and just over me. How a bright yellow pickup could be confused with a deer is beyond me.
Every single people working in their property should have one working non-stop in the deer season to signal that is not a deer. Some just are deertonic and unable to distinguish it from cars.
It wasn't. It was confused with a government vehicle. The one day a year that is synonymous with hunting accidents just provided plausible deniability for taking a few pot shots.
Basically, intensive agriculture is compensating for its destructive nature by just brutally suppressing any "undesirable" growth. In nature, pests target the weaker plants. Basically, you stress out plants by mismanaging them and they are fair game for a wide range of fungi, insects, etc. E.g. under or over watering them, depleting the soil, etc. all stress plants out. If you have unhealthy soil, it's hard to have healthy plants growing in it.
Intensive agriculture starts with disrupting the soil (by plowing it). This causes the soil to degrade over time and releases captured carbon into the atmosphere. The natural response is for aggressive weeds to start growing in disrupted soil. To "fix" this, farmers use a mix of pesticides and fertilizers. The pesticides kill a lot more than just the pests and weeds and then further degrades the soil and the bio diversity in and on it. And finally the fertilizer is just bleeding out of the land straight into the ground water and surrounding areas. It works but it is resource intensive. You need lots of water, chemicals, energy, etc to grow a single crop at the cost of literally anything else. Basically modern intensive farming grows mono cultures on top of artificially created deserts.
Organic farming avoids doing that. Try not to disrupt the soil. Kill off (and compost) weaker plants rather than trying to rescue them. Use nature to keep pests and weeds in check. Etc. It works but it is more labor intensive. You need people to look after things. You get healthier and tastier produce (cooks love organically grown stuff for this reason). However, automating that labor makes it more scalable. That's the opportunity here.
By what measure is vertical farming more efficient?. I've heard the contrary regarding green house gas emissions, it is still less efficient in most crops, it consumes too much power and that overcompensates any efficiency gained in other areas
It's interesting that this conclusion can change drastically if electricity becomes 10-100x more efficient
Vertical farms are essentially closed systems where nutrients, energy and some water goes in and produce comes out. There's a bit more to it of course but that's the general idea. They are climate controlled environments where light is provided via grow lights. In short, it uses energy which needs to come from somewhere.
You are confusing power requirements and greenhouse gases. Mostly these things use some form of sustainable power. Either sourced indirectly via some electricity provider; or directly with e.g. solar panels. If you think about it, the number one expensive consumable here is electricity and the cheapest way to produce it is via renewables. It's that simple. Why would you get more expensive energy if you can get it cheaper. Keeping all those ACs, lights, pumps, etc. going you are going to have to figure that one out or deal with competition that is smarter than you on that front.
There are probably a few farms that are forced to buy via the grid and are hopelessly dependent on their local coal burning monopolists. Not a great business plan but the good news is that coal is rapidly becoming a very unfashionable way to produce power due to the fact that it is expensive and dirty. It being expensive cuts down on energy margins and causes coal dependent producers to be dealing with awkward things like going bankrupt and being forced to pay more for their coal than they can charge for the resulting energy when wind is blowing and the sun is shining. Gas has similar issues, as people that pay gas bills these days can vouch for. Mine went up by 2x recently. Not great.
Solar's efficiency is below 50% [https://www.nrel.gov/pv/cell-efficiency.html]. Considering that, I'd guess for each m² of solar panel deployed, you'll get <50% of the yield that would be produced in that same land with horizontal farming (assuming fertile land).
You can argue that solar panels can be deployed in non-agricultural surfaces, and that's true, that allows to grow things where it was not possible before. But I'm not convinced that vertical farming is more efficient than horizontal on very important metrics like cost and GHG emissions
It doesn't even need to use herbicides because it removes weeds mechanically. We actually build it to also fit into Europe's small(er) scale farms as well.
Regarding the energy issues mentioned multiple times in various threads here: We managed to make it solar powered and battery buffered to work autonomously for more than a working day.
Disclosure: Co-funder of the company
But I'm dreaming, certainly if it would have existed humans would have tried to leverage them a long time ago.
It used to be our fav tools, now we just don't go in that direction much anymore.
We could study, teach, bread, terraform, integrate and so on at a much larger scale and with better tech. It's just... It doesn't look like a product.
You are missing something important: soil compaction is a sub linear function of weight. which means that larger tractors that cover more area with the same tire tracks are must friendly for the soil, compared to many small tractors. Because of this factor larger is better. Eventually we may get to the point where the tractor runs on tracks, to allow even bigger tractors that touch a smaller are of soil.
Sure flying drones will have less soil compaction, and they will have a place in the future, but keeping something in the air uses vastly more energy than wheels on the ground.
To cultivate the earth we're going to mine other parts of it? I'm not suggesting getting rid of the tools that improve quality of life but does automating one aspect not cause 100 other things down the line? There's mining, factories, repairing, replacing.
Maybe instead of robots we work on getting people to have gardens on their property, in cities use the rooftops, stop importing foods from across the globe, etc.
True, but the issue though is the problem they solve: famine.
How hard is the jump into the realm of computer vision/AI/robotics for someone with strong skills and experience building web applications? Is the coding much different? Is a company like this employing math/theroy-heavy people to develop new types of AI or are they working higher up the stack and just need people who can pipe data around?
If you don't want to change what you're doing or using, most modern robotic platforms have a web interface somewhere in them for monitoring the fleet. Any web developer would find their skills perfectly compatible.
For example, at Foxglove[0] we are building open source web-based visualization and data management for robotics (shameless plug: currently hiring).
[0] https://foxglove.dev/
https://www.fossjobs.net/ https://github.com/fossjobs/fossjobs/wiki/resources
If you are willing to make the jump. It's possible. You have an advantage.
> Is the coding much different?
Not as much as they have marketed you to believe.
> Is a company like this employing math/theroy-heavy people to develop new types of AI or are they working higher up the stack and just need people who can pipe data around?
No. They are not making academic thesis. You'd be surprised how primitive and "basic" many of these companies are (comparing to their marketing videos, which are click/eye baits). You can definitively make a difference there.
Of course, as other comments have mentioned, these companies still need front-end, databases, backend, and you know, normal basic automation scripts like everyone else.
I have worked for one of these companies that produces fancy AI based hardware.
Turns out the hardware is made in China and the AI is handled by third parties (amazon, etc.)
Does that mean putting it all together is trivial? Of course not. But it's not as difficult as it might seem from the outside.
AI-wise, nowadays model training stacks are practically codeless. Training a state of the art model on custom data usually boils down to changing some parameters in a config file, plus some boilerplate code. Not always that easy of course, but training an image classifier today can be a single line of Python.
It's not so much piping data around but dealing with datasets in any kind of non-trivial problem is where the hard work is. Neural nets are just another tool for most data scientists. For example exploring Kaggle will get you familiar with weird datasets, but actually collecting, arranging for data to be labelled, and then cleaning it is a skill on its own.
That's usually where experience comes in. Hence there are a lot of startups charging frankly obscene amounts of money for data labelling and curation tools.
So you won't actually have to jump into a whole new career in order to contribute to these types of products. You just have to search for a company that works on a challenge you want to contribute to.
Then again: Speaking from experience, most uses cases for ML are not fancy, world-saving new approaches like the one in the video but revolve around quite mundane applications and tasks like "how to make users click on this add", "how much will we sell next month", "can we automate the creation of thix Excel-file" etc., so pick your company carefully
We have a mixed team, including mechanical engineers, web developers, electronics guys, embedded software guys.
We do have ML people for custom networks but most of it is object detection/ classification which is a solved path and just needs data management.
We’ve done lots of cool ML applications, but many of them need management systems too so that’s where web apps come in.
You could definitely fill a position in a small company like mine and upskill / sideskill in time.
We’re based in Scotland but do have remote workers in other countries.
edit: forgot to include link :)
http://www.holovaty.com/writing/machine-learning-thoughts/