Hacker News new | past | comments | ask | show | jobs | submit login
Farming robot kills 100k weeds per hour with lasers (freethink.com)
684 points by HiroProtagonist on April 26, 2021 | hide | past | favorite | 464 comments

Given where ML and CV are today, I'd bet on open source models trained on weeds within 3-5 years. This company can scale, but their unit price is going to plummet. Also, it sounds like they're using cameras, where it's concievable that other future sensors could be more efficient. An iteration of this with higher resolution cameras and small flying drones seems like an intuitive next step.

This is a super interesting problem because the confusion matrix (fp/fn/tp/tn) rate that makes this economical is going to be variable across both crops, and market demand.

If there suddenly there was a demand jump for peas, you could afford to use a model with less accuracy, because you are optimizing betwee a sunk labor cost and margin on your yield. You could literally tune your detection parameters based on futures price data, since if if prices were high, you could optimize compute on your model. Anyway, spoken as a total outsider, but what a cool and interesting set of problems.

Drones do not seem like a logical step for me. I'm not a laser expert, but seems like the energy storage needed to kill hundreds to thousands of something with light is heavier than flying would realistically allow. Wheels and tracks are what farming is built around, no need to literally reinvent the wheel here.

Drones also have poor flight time to recharging time ratio. Which means you need more of them to get the job done.

Farms are designed to be serviced by farm vehicles. A vehicle makes a lot of sense.

If they were able to make drones work, it could open up more variable farm designs: rows are mostly necessary so that farms can be serviced by vehicles.

Polyculture farming could become much more economically feasible if drones could weed out all non-whitelisted species.

Would also be a great boon to forestry: would be awesome to make a bunch of drones to fly through forests and zap any non-native invasive species it sees.

>If they were able to make drones work, it could open up more variable farm designs: rows are mostly necessary so that farms can be serviced by vehicles.

Sure, but the weedkiller isn't the only vehicle that needs to work on the farm. I don't think flying drones are going to be ploughing any time soon.

> If they were able to make drones work

Yes, let's put powerful lasers on drones flown by AI. What could happen? I just hope the human recognition model doesn't shoot more of one race than another.

> zap any non-native invasive species it sees

Given the climate change it could be the other way around. Zap some native saplings, which can no longer thrive.

I'll bet there's creative solutions that could take advantage of a hub and spoke model. Picture a solar recharging base station that drones can dock with to charge while others rotate in.

Drones don't need to be battery operated - but, I agree that there is no reason to use them in fields which are already designed for vehicle access with semi-standardized dimensions.

If they're fuel powered drones, they need to refuel. Plus that's potentially expensive to operate depending how many you need and how heavy they are.

There's probably something to be said for hard-wired drones here. The weight of the cable is something you have to contend with, but with a physical wire you can run a larger drone longer.

How is that better than a wheeled drone dragging a giant orange extension cord behind it?

I'm not sure that's a great idea either, but flying brings more problems than it solves in this problem space.

Given that the plants are on the ground, I imagine not sweeping over them with an extension cable would be an obvious advantage. Still, you could work with suspended powerlines as easily for the big-box drone as with tractors, which brings me to another (probably dumb) question: Why has nobody built a landline electric tractor?

Farmers sweep things along the field all the time. The standard farming sprinkler is supplied by pipes / hoses which are bridged above crop height between wheeled towers. As the spinkler works its way in a circle on the field around the water supply, the towers roll over a negligible amount of crop.

> landline electric tractor

Well, there are some robots that travel on aerial lines, typically used for cable inspection. Here's a quick search result: https://spectrum.ieee.org/automaton/robotics/industrial-robo...

For vehicles, crops yes, animal husbandry often roads get you to parts of a property but are far from covering a property.

Helium assisted drones could provide longer air time.

Hydrogen. Even lighter, much cheaper, and would save on a (very!) finite natural resource.

And, hey, with unmanned drones over rural fields you won't get any Hindenburg disasters.

Maybe very small drones with mirrors and cameras and a ground-based robotic tank with a laser cannon.

Or it could just fly upside down. maybe add wheels to set cutting height :)

The scene from Spies Like Us always comes to mind when bouncing lasers around. Or the Real Genius as well. So you better be careful using this around corn fields, or we know what can happen.

I love the image of this in my head, but I can imagine in practice that many farms are not on flat terrain where there's direct line of sight [fire] for a central station to flash a lazer pulse. Perhaps if you had many smaller stations with these laser cannons spread about the field and each cannon would coordinate with the drone in the closest line of sight.

Pretty sure they're explicitly suggesting indirect fire, actually. Base station fires at mirror, and drone uses mirror to reflect laser to target. There might be stability problems, angle-at-which-it-hits-the-weed problems, and more-laser-scattering-due-to-longer-beam-path problems, but line of sight itself should be fine since the drone is up in the air.

Heat-ray welding Tripods.

This is an incredibly great idea.

Indeed a good idea, but I would only hope it would have a reliable fail-safe, becaue it would be a little reverse GoldenEye.

Spoken like a true software engineer.

To me that message implied weed recognition and marking with drones followed by destruction with the cube.

Ah, the classic "Death Cube and Tracker Drone" design pattern. DCTD for short.

That's my understanding too but what about precision?

Both the drone and the ground robot would need milimiter precision geolocation to coordinate otherwise the laser can miss the weed.

No they don't. They need fairly coarse tracking (A few feet or so.) for geolocation as that is only to make sure you end up covering the whole field.

Now for individual weeds you need precise aiming but that's only from a few feet away at max.

The overall geo location boils down to being able to track a row as you go down it and then go to the next row at the other end. The lasers don't care about the geo location at all.

Unless you have a precise map of all the weed locations, precise location in absolute world coordinates doesn't seem too important. You'd need fairly precise and accurate relative positioning from tank to drone and drone to weed, though. Drone to weed could be done via camera (you need to detect the weed anyway), but tank to drone would likely be a difficult engineering problem.

A flying drone seems like the wrong way to solve the problem, though. For a drone close enough to the ground to reliably detect weeds, you'd likely need a multi-rotor, a slow flying fixed wing aircraft, or a blimp. A multirotor has a huge energy penalty, and a fixed wing aircraft or blimp loses the practical gains vs. just using a tractor of some sort.

RTK GPS has +/- 3cm Lat/long precision. That gets you very close, and then with a confirming camera, you could aim and fire.

Completely agree about flying: but I wonder about the wheel.

A spider-carriage walking robot which could step around the valuable plants, climb steeper hills, and wouldn't dictate a row-and-plow approach to agriculture, that could be pretty compelling. Less soil compaction, more flexibility, could work inside forested regions as well. It has potential.

IMO; I think setting up lightweight aluminum rails, and run a framework of multi-purpose bots along that. Since the technology is getting lighter and lighter, it seems you could even do this and use LESS material than these gargantuan tractor/harvester/combine units. It may not even be necessary to plough fields anymore, when a smaller automated unit could just operate on the least amount of soil necessary for the plant to thrive.

Also: no more chemical weed killers.

Do this with bugs, and get rid of insecticides as well.

It's not just the robot though, it's the millions or tens of millions of dollars of wheeled or tracked equipment he already has.

Still easier to harvest if you have the rows though.

Good point, but you havent thought of the *millions of tiny mirrors* they'd obviously be using :)

One mirror mounted on the drone would be enough. And a strong enough laser device at a stationary position. Of course, the challenge would be to hit the mirror and not the drone ...

They'd probably just use rare-earth magnets.

DJI has some demonstration of target pesticide spreading drone https://www.youtube.com/watch?v=mfdWYztSqUI It's not as targeted as laser, and it has residual problem, and requires the crop to have resistance, or the pesticide/weedkiller to be specifically targeted.

But I think it might be possible to focus sunlight to generate energy to kill weeds, not through battery or generated electricity.

Drones will be a big player in next 2-10 years (search for Rantizo). The more they can operate autonomously the better. The small payload size on drones aren't a huge problem if you can have 'refilling stations'

agricultural drones run on gasoline and have plenty of energy

gasoline? I'm thinking ethanol derived from fermentation of the parts of the plant that were not shipped to market.

yeah could be that as well, definitely some conventional ICE engines, not batteries

When they can scale this down for home use, I think the market will want legs. A miniature Imperial Walker ambling around your garden zapping weeds would sell.

What about putting them on an electrified wire? String a grid of them over the fields for motion and for power. Power the grid with solar.

Any kind of wired setup would interfere with other tractors they use in the field. Fertilization, and eventually harvesting.

I’m not sure why people are inventing more complicated solutions when this robot seems to handle the job quite well without them. It covers 20 acres in a day, you can likely keep 100+ acres weed free continuously without any other special gimmicks.

I think most people here don't run large farms, and actually want something for their garden.

It stands to reason that if you're going to be running wires over the entire field you wouldn't need flying drones so much as you would need cameras on wheels (like they have at football stadiums).

If the camera sees the correct spectra for a known weed it can drop down and get a closer inspection then mark the spot or just burn the weeds with an attached laser.

I assume the crawler is necessary because--in order for the laser to work--the weeds need to be identified very early as they emerge from the ground. You might not be able to spot them at such an early stage from above without expensive optics.

Yes totally agree on the benefits of early detection — easier to identify eg. once the desired crop is larger than, say 10mm, then just fire the laser on anything green that is less than 2mm, adjust those parameters as required as the crop matures, also getting them early means less laser energy required to kill it, longer battery life, larger acreage covered

Wire is too expensive for that. You need thick wire to hold itself up without posts every few meters, which in turn means it is more expensive than the normal house wires. Not that it couldn't be done, but it is too expensive.

Now there is opportunity to have one long wire that the tractor reels in/out as it makes passes. This isn't a flying drone, but that isn't really needed for anything other than cool.

No, wire gantry would be very much ok for this solution.

You don't need wires to be that rigid for that.

I would see the drones being used to help map weed concentration and optimize the route/efficiency of the weed killing robot, not using drones to kill the weeds.

There are already solutions that use drone photogrammetry to map crop health, ground coverage and so on. It feels like a logical next step to use a drone to assist mapping the best route/find problem areas to target for the weed-roomba.

>I would see the drones being used to help map weed concentration and optimize the route/efficiency of the weed killing robot, not using drones to kill the weeds.

That's a ridiculous solution to a non-existent problem. The robot literally has no rush to go to the weeds. It can crawl along the field by itself 24/7. Throw in a solar panel charging station and you literally have free energy for it to piss away.

Instead you want to add complexity to the setup, increase maintenance costs and potentially shorten the lifespan of the system.

It’s not ridiculous at all, hours count. I haven’t looked at the article, but I’m assuming it’s the same machine, it’s a 10,000 pound unit with tiny little tires and a 75 hp diesel engine. It’s going to compact the soil, get stuck, and waste fuel driving around looking for weeds.

If sending a 20 pound gas drone with a 2 TB solid-state drive and 60 FPS 4K camera on it up and down the field for one 100th of the fuel consumption once a week saves 500 hours a year off that beast, it’d probably be worth it.

> It’s going to compact the soil, get stuck, and waste fuel driving around looking for weeds.

This true of any modern farming equipment, which is why planting strategy is so critical--I kept messing up in my first year as I was going way too fast instead of taking my time and my planted rows were never straight, it was in a green house and we had an old diesel tractor so my lungs were hurting after the first hour and I just compromised on that aspect.

But then when I had to go back and weed, maintain and eventually harvest the oddly planted fields of salads and potatoes I made sure to follow the natural ebb and flow of the soil compaction, which was there even after it was tilled.

What I'm saying is that the compacted soil can help you in later seasons to maintain the direction of growth so it's not entirely a bad thing to have. And unless you farmed the same fields for several seasons/years you would overlook this as a net benefit. Eventually seasoned farm hands can do it while drinking and smoking and just listening to how the engine is struggling without even touching the steering wheel, they can even get out of the tractor and walk ahead of it as it's crawling forward to check if everything is fine. Where as for apprentices like myself who had no real experience in Ag it was a remarkable discovery that one shouldn't overlook.

Also, drones are already widely in use; they monitor temperature and moisture in real time and a offer other services right now [0]. Their is a company that I found that posted here on YC a few years ago that offers this as a service [1].

I was actually going to get my commercial drone's license a couple month's before I retired as a farm manager in Hawaii due to extensive injuries, but my fintech startup required saving and scraping all my pennies to afford getting it off the ground. I might consider getting it again as things start to be more widely deployed and the costs keep coming down and I still would like to be involved in some capacity in Ag.



That does depend on how much area it can cover per day.

If it covers all your land in under a day, sure. If not, then route planning may mean you can have one rather than two or more of these very expensive machines.

It's not a drasticly complex addition, and mapping weeds with drones appears to be a use of them already.

> That does depend on how much area it can cover per day.

This assumes weeding is a time sensitive task, which I find very doubtful. Realistically covering your land area in a week is probably enough, but maybe that's much higher.

Ok thank you. Common sense on non-programming topics on HN is more rare than I would like. You clearly understand the problem better than most

Tethered drones might work.

I don't think you really need AI/CV if you have multispectral or full-spectrum imaging. Different species will have different absorption/emission at multiple wavelength. This makes it easy to identify crops, and seems to be commonly used for aerial imaging[1-3].

That kind of sensor is expensive though, and while you could probably do it for cheap with something like a DLP wheel (edit: or an array of different light emitters) plus a B/W camera, ML might be more price-effective, though probably more error-prone, so it isn't a given if you want a high match rate.

Also, isn't this a fire hazard?

[1] https://en.wikipedia.org/wiki/Hyperspectral_imaging#Agricult...

[2] https://en.wikipedia.org/wiki/Multispectral_image

[3] https://en.wikipedia.org/wiki/Full-spectrum_photography

> This makes it easy to identify crops

...until they evolve to emit the same spectrum of light as the crops being harvested! Cuz you know that's going to happen.

Well, that's right, and I didn't think of it. But besides what the sibling comments pointed out, that's an issue with any kind of weeding, including manual.

The answer is to select according to a wide range of criteria, and not rely on a single one. That way, weeds cannot progressively acquire resistance, and need to check all the boxes at once, which makes it highly unlikely that they will pass on their "slightly better" genes.

And of course, the larger the scale you employ a single weeding system at, the more risky it gets. It would be great for supplementing herbicide (while lowering doses) at first, for instance.

This is the part where Monsanto starts engineering crops to fluoresce in a particular frequency of light, so you can target weeds inversely.

That makes me think that this also makes crop-rotation obsolete. You can plant 20 different crops in the same field, in repeating sequence. The bot will take care of them as-needed, and next season, move everything down one slot.

This is a known process already https://en.wikipedia.org/wiki/Vavilovian_mimicry

Examples of this are Rye and Oats.

That’s a slow process, further you can always rotate between crops with different spectra.

It sounds like a slow process but plant life cycles are yearly compared to 20+ years for human generations to spawn, it will happen faster than you think. There are already round up resistant weeds and it's only been 50 years.

Even 30 years after first implementation would be several generations of better hardware and software. Spectra is just one way to get v1 out the door, there are many ways you can improve things.

Crops are planted in a specific pattern, at a specific time, grow at a specific rate, and have a unique shape. So you have a lot of information to work with.

Bacteria and viruses do that, but weeds first don't reproduce as fast and second humans can see them and find out something like treating crop seeds with fluorescent color.

Eventually, the only organisms that remain will be mutualistic ones, if we do it right.

We just need the weedbots to test plants for flavour and nutritional value; cf. xkcd.com/810

I think the AI solution is actually pretty well proven. I use PlantNet to identify weeds in my yard and it's plenty accurate and doesn't require any sort of advanced sensor.

Could be, but having more discriminating data as an input can only improve speed and accuracy. I don't know what technology they picked for the first iteration, but can only imagine they'll end up using every trick in the book going forward.

A slightly more advanced sensor or lighting apparatus wouldn't cost much more, proportionately to the cost of the whole system.

The cost would be acquiring a training data set that includes data from the more advanced sensor, versus existing catalogs with many millions of images and tens of thousands of species.

> Also, isn't this a fire hazard?

Unlikely. Heating a plant to kill it is considerably lower power than heating a plant to combust it and turn it to ash. It would also be trivial to add a water sprayer as part of the process.

> you have multispectral or full-spectrum imaging

...that's exactly what CV needs to handle.

Well, right, but it doesn't need particularly impressive algotithms to achieve something with that data. Those algorithms were here 20 years ago, with less processing power and camera resolution.

A lot of modern ML is not that impressive either.

I can imagine discriminating species of a hundred leaves in a very high resolution images to be very challenging.

The leaves are 3-dimensional objects can be twisted, tangled, curled up, broken, half rotten, sunburnt...

Actually, the next most logical version is an implement that runs off the PTO drive of a tractor, not drones. One needs to piggyback off existing infrastructure, not replace it.

My high school summer job actually was helping with wheat harvest. You can only work when it's hot as you get better prices on the harvest.

There were certain kinds of weeds that really clog the combine (usually ragweed, not a huge problem in the early summer, but weird things happen) or cause damage to the cutting head (invasive brush/tree species). We walked the field and pulled anything really bad out.

Talk about a way to earn $25/day. You only finished when the equipment was put up for the night.

It doesn’t need to recognize every possible weed species, just “crop of interest / not crop of interest”. A much simpler but still valuable task.

That's what made sense for my company; we just identify the crop, and then churn up all the areas of dirt that are "not crop". Also, hey, free cultivation. Do it often enough, and your weeds never even get that large, which also greatly increases your identification accuracy.

A real world useful variant of Not Hotdog. https://news.ycombinator.com/item?id=14636228

That was my first thought as well.

My second thought was, "Please nobody invent a drone that shoots lasers at 'not hotdog'".

I wonder why it can't it be done via location? Where we planted, vs where we did not plant.

This is the comment I came here for.

I had a professor in college who was building self-driving tractors and would come in every other week complaining about John Deere this, or Case that, trying to steal his business with more expensive solutions. It turns out you can use GPS for a rough location and a fancy $200 gyroscope for millimeter precision. Then just plant the seeds on an exact grid and you know that anything not on the grid is a weed.

And actually, his suggestion was to use high-pressure water jets to cut the weeds instead of lasers. It would/could be less energy-intensive.

RTK GPS is used as a second factor to vision in these machines. It’s just not good enough to target with as a sole / primary factor in the real world though. Bit of drift and whoops, $30,000 of crop gone.

Lasers are not used because they’re expensive and dangerous. And Co2 lasers (as per the machine in the article) are powerful but super fragile.

Water shooting around at high pressure is in no way efficient or easy to handle.

Compressed light is the technology that’s actually going to be used for precision weeding. It has the speed, power and simplicity of lasers, without the cost and danger: https://www.linkedin.com/feed/update/activity:67899797724715...

Its like cabling or rails have never been used before to guide a machine along its intended path. Alternatively, markers could be placed at planting.

Thanks for posting this, very interesting. The linkedin profile pointed to a youtube video (unlisted) that was pretty interesting and covered how they could calculate eye safely using simple math that indicated the safe distance was 2 meters away. They appeared to be using blue LEDs at high intensity to char-or-inactivate weed photosynthesis.


it'll probably end up being something like a gamma ray source; as probably the cheapest and most durable solution.

Hot dog, no hot dog?

My prediction, this thing will be built into pivots.

Most farms irrigate with a giant automated system that sweeps around the field in regular intervals. These weed death lasers could be integrated to the base of the system and ran from the same power that moves the pivot.

Since these things don't move all that fast, you could have a central track and move the killer box up and down the line.

It wouldn’t even have to burn all the weeds on the first pass. It’s like the herbicidal equivalent of mark-and-sweep garbage collection.

yes, this seems like a natural progression.

> Also, it sounds like they're using cameras, where it's concievable that other future sensors could be more efficient.

Bit of a pedantic note:

What other sensor could you use?

I'm grasping at straws (pun not intended) to figure out any other modality that could work as well in a giant field of psuedo-randomly mixed plants with the wind blowing chemical signatures all about.

I assume they mean "visible spectrum" when they say cameras, because that's what most off-the-shelf systems are tuned to.

A spectrometer is a single pixel camera, I guess, but it isn't being used with imaging optics, and it isn't being used to stitch together a photograph.

In a general case you could embed some version of GFP instead of glyphosate resistance. Then you can set the system loose (within the field, lol) to actively interrogate plants, zapping intruders that fail their scans.

Don't think there will be a need to bio-tag weed versus no weed. Current tech and sensors can probably get to 2-3 9s worth of accuracy (bonus, you don't care about accidently hitting/killing a false positive or two, there are many plants if this is a commodity crop)

I wonder if Monsanto seeds will start producing a signature that could be recognized. This will help in 2 ways. First, more easily identify those pesky farmers using their seeds without proper licensing. Second, help the machines know what plants to keep, and the ones to remove.

Monsanto.....doesn't exist anymore....Laughable people hate on Monsanto/Bayer for protecting their IP, but when big tech does it it's just good business

But if you add a bio-tag, you can make your system look like ED-209 and recreate this scene: https://www.youtube.com/watch?v=iYUnp-x0SDE

I think the question was just visible light cameras vs other wavelengths.

With visible light, all you have to go on is small differences in shades of green and shape of the leaves. With other spectra you'd have more cues.

i'm sure i'm oversimplifying but if you can reliably identify your crop then everything else is a weed. Maybe put some kind of GMO marker that makes what you want to keep stick out like a sore thumb then you just nuke everything else.

If you know where your crops are, everything else is a weed.

This is the traditional approach and it really blows my mind that so many hammer-syndrome AI/ML/flavor-of-the-month VC spenders haven't simply gone to see for themselves what works on farms. It probably doesn't even register to the casual observer, but planting a careful row of tomatoes with the root ball in a particular direction is a fit for the cultivator tool to come by and turn the soil up on the stem of the young, but now established plant.

That's actually the "weed control at scale" developed in conjunction with the tractor.

I'm not a fan of modern conventional agriculture. The abuse done to topsoil is terrible, and we need better systems. But new systems need to keep their eye on the ball and the ball is a John Deere pulling a 40-foot cultivator across a field while the "operator" reads twitter ( or <verb>s Clubhouse ) only looking up to mind the turns.

Everything starts somewhere, but just because your tech has ML and Laserbeams doesn't mean it passes the tool/toy test.

Perhaps he just meant IR sensors? Also, I wonder whether UV sensors would be helpful.

I wonder at what point it becomes economically feasible to breed crops that are more distinguishable from weeds?

After all, this is what we did with chemical weedkillers.

I wondered about the corollary - at what point does selective pressure create a weed that's indistinguishable from a crop?

I doubt weeds will be able to maintain advantages that let them compete well to begin with, while also evolving to evade all of a series of detection techniques that will be added as they evolve their way around the first few. To stick around they need to survive well outside of fields, too, and I'd expect detection-resistant varieties to become increasingly inefficient at surviving in areas where they're not being lasered to death.

well, same, do we have weeds that have evolved resistance to roundup?

Commodity drown with high powered laser for killing things based on image recognition sounds like sci-fi dystopia.

True, but it seems less dystopian than soaking vast fields with millions of pounds of neurotoxins, like we do today.


Hey, I get it, needles are scary. You don't need to make stuff up to justify a completely rational fear of getting jabbed with a needle.

I'm not concerned about the needle, only what's coming through it, approved by an entirely captured regulatory body.

can't be any worse than what's potentially in your hamburger.

Then you drink a cup of glyphosate. Even folks on payroll spewing that nonsense would never entertain the thought of actually going through with it.

There is a video on youtube of a Monsanto lobbyist being challenged to do exactly that, quite funny.


Yeah, I imagine massive overuse on "pest" species. Mosquito eradication will be a high priority for such technology, but there will be people who modify it to eradicate everything that flies or crawls. Everything smaller than a cat, if they could.

I think lasers for killing large animals (including us) will be prohibitively expensive for a long time to come, fingers crossed.

Don't worry. The robots can use machine guns like the rest of us.

Unfortunately, you don't need that much power to permanently blind a human.

for that matter; microwaves cook us from the inside out at 100+ meters, and don't require very good aim.

Bill Gates was already involved in a startup to radar identify and kill mosquitos with lasers . . . guess it did not go well.

It resulted in a prototype that apparently cost $50 per unit, but no retail product yet.


Are there massive amounts of retail products in, say, Africa?

Yes, Africa's retail market was worth US$1.4tn in 2015.[1]

1. https://www.brookings.edu/wp-content/uploads/2018/12/africas...

I know a lot of farmers that would respond very well to having "laser weed killer" as a product. I get the feeling the video feed would be amazing. Throw in pest killing and you own the market.

Although I keep thinking the movie Runaway (1984) is going to be a much truer representation of reality than I would like.

Just don't step on the field and you won't be mistaken for a rat and be zapped.

All roads lead to Skynet, some less obvious than others.

I'm not sure that "killer robots with lasers that can be field-trained to identify specific targets" would be quite the blessing that everyone seems to think.

Why do your weeds have 2 eyes, a nose, and a mouth in all of the training libraries?

>An iteration of this with higher resolution cameras and small flying drones seems like an intuitive next step.

The gimbal alone on your flying drone will cost more than the entire river.

Non-engineer speculation. Seems like drones might have the potential to make it cheaper? What if the laser was tower mounted and the drones just were there as a camera and reflector and to verify clear line of sight? I guess that would only really work well when the plants are small...

My first thought was that ML-guided flying drones with laser weapons seems like it could be catastrophic is hacked. Imagine someone changed the drones to recognize humans (or, assuming the lasers can't hurt a human normally, the eyes of automobile drivers) as valid targets.

Open Source models don't need to be trained on weeds. They need to be trained on whatever product is being grown. Everything else can be zapped regardless of what kind of weed it is.

Sounds a lot safer to whitelist what can be zapped rather then blacklist what can't, no? I mean, this is a moving autonomous laser platform we're talking about.

Microwave powered drones. Have a mothership or tower track and send energy to the drones.

I'd like a lawnmower sized one for my yard!

I want to see details on power source and energy usage. That's where I'm a bit skeptical. To do the 20 acres they're talking about you'd need one hell of a huge battery, assuming it's electric. Even just to move the thing. Then add power draw for the laser on as well.

Also seems odd to have it be something that is self-powered and autonomous rather than just something you pull behind a tractor on the three point hitch (where you could power off the PTO of the tractor). Seems excessively novel, given most farmers are already spending time going up and down the rows cultivating, etc. anyways. Adds to the cost, and complexity.

>Also seems odd to have it be something that is self-powered and autonomous rather than just something you pull behind a tractor on the three point hitch

That doesn't seem odd to me at all. The whole point of this thing is to let it run all day killing weeds while you do other stuff. The fact it only has a max of 5mph (and likely significantly slower when there are actual weeds to kill) means you would spend all day every day in the field trying to take care of weeds.

When they spray for weeds they're going significantly faster than that and cover a massive swath in one pass, and that's generally outsourced to someone other than the farmer himself. This looks like it's good for maybe 3 rows at a time.

> The whole point of this thing is to let it run all day killing weeds while you do other stuff.

There's a lot of automated tractors now

I guess I don't follow what the advantage would be. Putting hours and wear/tear on a $250k tractor that you likely have other uses for in order to cut what? $10k off the price of this for the cummins motor and hydraulics? It would need power generation of some sort, so you'd be doing PTO off the tractor instead to drive a generator.

I'd imagine the end result would be more fuel and in the long run more expense.

The advantage with tractor-pulled is lower cost and higher reliability. The propulsion, guidance, and power-supply problem has been dealt with by industry already and is already owned by the farmer. The autonomous weeding machine will undoubtedly have issues with these three functions, which means down-time and cost. The machine would be cheaper if tractor-pulled, and the developers could focus on the problem at hand. Also, the tractor driver could periodically stop to monitor and tweak the weed killer, especially important given the new technology. We did a recent project with a farming implement that could have been autonomous, and when one is realistic with reliability and maintenance costs (unless one intends a McDonald's ice-cream machine repair business model) then the argument for tractor-pulled is very strong.

>The advantage with tractor-pulled is lower cost and higher reliability.

Based on what? The cummins engine they're using is bulletproof and a rounding error in the cost of the unit. Hydraulic motors will run for 10s of thousands of hours without any maintenance beyond a fluid change.

People in this thread keep claiming tractors are fully autonomous, which model? If they aren't fully autonomous, what farmer is volunteering to spend hundreds to thousands of hours in their tractor doing nothing but putting along at 5mph stop-and-go while this thing zaps weeds?


My theory is JD,Case already have the tech to be level 4 autonomous but they haven't figured out how to make more revenue than selling $250K quad-tracs with 'some' autonomy.


Thousands... millions of farmers already putting along doing nothing but dragging a cultivator / sprayer / fertilizer spreader / bush hog / rototiller / weed badger / manure spreader etc. etc.

I assume you haven't farmed? You don't run any of those implements at 5mph with a 3-row spread. It would literally be impossible to run a modern farm at those speeds and spread.

100% agree and that's what my original comment was trying to get at.

I can see selling two models, one autonomous, one three point hitch / PTO. The advantage of the latter is clear to me.

But then the product starts to look a lot less sci-fi, doesn't it? We already have pull-behind weed burners that use propane torches and not lasers. The only "magic" would be in the AI recognition systems (which I have questions about.) Perhaps one could not get investment $$ for it then :-(

Also propane torches seem more efficient to me than converting diesel combustion to electricity to heat energy.

I think the difference is that propane torches are an imprecise mechanism generally used to clear dead area between planted rows, while this laser based solution could be used selectively within a planted row (provided it is real at all).

Fully agreed about the 3-point comment though. Why take on building an autonomous tractor AND a targeted weed-killer, rather than tackling the differentiating problem only? Seems like a hype train measure. Or the systems integration is very important, in which case a partnership would be the obvious route or white-labeling an autonomous tractor. Regardless, this strategy seems very weird to me too.

Right, which is why I think it's likely a result of VC direction. It's not enough to have a profitable or sensible product, it has to be something that can sell in acquisition for 10 or 100 x the investment given.

"Autonomous vehicles are hot and AI is hot, go with that."

Yes, but autonomous vehicles with AI don't scale at zero cost like web apps, so the valuation shouldn't have the same multiplier.

Yeah. But I wonder if putting it all in one was for better coupling between shooting laser control vs the whole cart moving around (for example, slowing down/stopping when the weed density goes up).

I noticed this thing uses hydraulic drive motors... I assume that was so they could run the engine as 'electrical first', but I also wonder if it gives them better start/stop control of the cart.

Many tractors these days are hydraulic as well.

"74-hp Cummins diesel QSF2.8"


There's a video on YouTube of a walk around. https://youtu.be/fK3AQgt47z4 The CEO of the company describes it as "an all diesel hydraulic system." Essentially there's a big generator to drive it around and provide power to the pew pew lasers.


Lasers are 150 watts, so a small fraction of the power draw.

Seems like getting the automation is key, replacing herbicides would require multiple passes over the same ground.

Just a point of clarification, for a CO2 laser, getting 150 optical watts out the aperture usually means putting about 1500 electrical watts into the tube. The rest comes out as heat in the water jacket, which you pass through a radiator or possibly an active chiller, and those pumps and fans draw their own power as well.

And before those 1500W get into the tube, they go through a high-voltage power supply, which is maybe 80% efficient if you're lucky. All-up power draw on a laser of that scale is somewhere north of 2KW from the wall.

(And before anyone asks: Yes this is very good efficiency for a laser, which is one reason why CO2 is so popular. YAG and fiber lasers tend to be in the low single-digit percents.)

150W of output power.

Lasers are notoriously inefficient, CO2 ones are in the 5-20% range.

So probably at least 1000W for the lasers, assuming 10% efficiency.

The modest engine powering it is 55 kW.

People who work in pulsed kW laser labs are all sitting here salivating at thought and also horrified at the thought.

Salivating because that's a lot of power that you can use for experiments, much more than most laser technicians have ever seen.

Horrified because not only are you going to go instantly blind (as usual), but now the rest of your head will smolder too.

55kW at 240V is only 230A. A typical house gets 200A service. It’s a lot of power, but you wouldn’t have to move mountains to get it if you worked at a lab that needed it. It could be had for less than the price of many used cars.

Maybe it's because the robot needs different amounts of time for different patches of soil? As in, it kinda needs to go its own pace to a) ensure it has enough time to detect and kill all weeds (which depends on weed density, presumably - the video shows it only killing one at a time) and b) maybe it has to stop to ensure it hits the right thing because it can't handle moving targets (not clear from the video)?

No idea if this is something that modern tractors could accommodate already or whether it would need some annoying human-in-the-loop stop and go.

Farming equipment is typically fuel based. And this appears to be the case for this robot as well: it uses Diesel.

At that point, is this more environmental than applying Round-Up to the field?

Each of these lumbering beasts goes 5 MPH, burning diesel the entire time. How many gallons of gas is going to be spent weeding a hundred acres of farmland?

How do you think the roundup is applied in the first place?

Well they are quite different pollutants. Hard to compare directly.

I know, and the two are often in conflict.

For example, replacing a deforested area might sometimes be faster by introducing a non-native invasive species. From a carbon perspective it would be a net-good, from an ecosystem perspective it would be a net-bad. Which is more important?

The QSF2.8 Diesel is a Tier 4 Diesel engine, so it is MUCH cleaner burning that most tractors out there.

Diesel engines can be eventually replaced with overhead wires

In a field? Not likely.

This is in no small part because of the massive subsidization of fuel for farming in most countries. Not only is the fuel for agriculture heavily discounted, but what you do pay can be written off against your revenues.

Same with most fisheries.

If governments want these sectors to go green, that is going to be a very big and painful band-aid to peel off.

You do realize that deducting the cost of producing a product is entirely normal and expected behavior right?

It’s not some special loophole.

If you use $1000 of electricity to harvest your crops, that $1000 would be equally as deductible because it’s just a cost of production?

> but what you do pay can be written off against your revenue

How is that different from any other business expense in any other industry.

Operating expense is different from a capital expenditure.

I don't see why power would be an issue. Why do you even assume there's a huge battery instead of a diesel engine and generator?

Having this thing be autonomous makes it more expensive to acquire, yes, but way cheaper to operate because the labor cost of pulling this thing with a non-autonomous tractor is quite large (even if the tractor were autonomous, having two autonomous robots doing different things is better than having one doing two different things that might halve its availability for each kind of task). This is a very big deal. If labor were a non-issue we'd have people weeding manually and we'd not use herbicides. Everything in farming is about labor, which is why we've gone from being agrarian societies to industrial and post-industrial ones: by bringing economies of scale to agriculture in order to greatly reduce labor costs in agriculture.

Why would you assume it was electric?

The energy usage should be minimal - it doesn't take much power to burn a leaf.

If you pull it behind a tractor, you need to hire someone to drive the tractor. This defeats the purpose of an autonomous weeding system.

> The energy usage should be minimal - it doesn't take much power to burn a leaf.

This is certainly true for dry plants. If they're weeds, they're actively growing, so they are trapping a lot of moisture that could make it hard to burn. If the weeds are a problem, then they're going to be growing. Many weeds, too, will continue to grow, even if their leaves are damaged or removed.

I think we might be using different definitions of "burn." I mean burnt by radiation so it no longer photosynthesizes -- in some plants this can happen simply from overexposure to sunlight. It's not necessary to turn it to ash.

A machine continually damaging cotyledons in emerging seedlings would go a long way. That's part of the advantage of a continually running weed killer.

Instead of a laser, I wonder if it would make sense to use positioned lenses and have the robot focus the sun's rays on each weed to burn it out. Though that probably would cause fires and would only be usable during certain hours of the day.

Why would the fire danger be different than the laser?

I could easily be wrong on this, but my understanding is that a laser is highly focused to one spot. A lens would have a wider spread and would also heat up the area surrounding the target to the point that it could ignite.

> would only be usable during certain hours of the day

Limiting usage to when the light will be strong enough to be focused would seriously limit the hours of usage

Maybe adding solar panels to harness solar energy to offset ICE emissions would be a consideration

That woukd be a good idea, though wouldn't that notably increase the cost of the unit as well?

The video looks like a mix of cgi graphics with unrelated real videos, so I guess they don't have a real machine yet.


(Video of an operating unit)

There is nothing particularly unbelievable about the device, and it seems, as such innovations tend to be, quite obvious.

It's driving on asphalt in a parking lot. Do they have a video of the device following the grooves over mud? (It would be not surprising, because other automated tractors can do that.)

It would be more interesting to see a video where they fire the lasers. Can it aim while moving? Can it aim while the engine is on and everything is vibrating?

"I want to see details on power source and energy usage. That's where I'm a bit skeptical."

Meh. Worse case you could even have a wire.

Hell, you could probably stick a fuel tank and a generator on it if you had to.

Which is basically what they did; upthread, someone links to the specs and they have an onboard Cummins diesel generator.

Check the Features & Specs page [1]

    74-hp Cummins diesel QSF2.8
    4 hydraulic drive motors
    75-gallon fuel capacity
I'd bet that diesel also powers a generator to run the computers and 150W lasers.

[1] https://carbonrobotics.com/features

You’d be dragging that wire over your rows of crops, potentially damaging them.

Oh it's for sure an energy hog and the 20 acres is marketing.

Super interesting especially think of it as in its infancy. Also would be interested to see how the real world results are. I'm guessing it can laser all life but I would be curious to see how well the algorithm worked with protecting certain species.

Love seeing advances in this space. I have an 80 acre farm and I’m waiting for an autonomous slope mower.

Question about this tech is, does it really kill weeds, or just defoliate them? Like, is this laser strong enough to get down to the roots, which are often very deep with weeds?

If not, I’m wondering why lasers instead of something mechanical that can corkscrew down and tear up the entire root?

Depends on the laser(s) and the weeds in question. Deep rooted weeds require different IR/UV than broadleaf.

They actually kill them because they cook the root. That's ALSO why it has to be autonomous - it can't be done at a run.

We worked on this when I was at MTD, but I think they abandoned the project after I left.

Wouldn't a microwave be easier (power, speed, penetration depth, etc)?

Depends on the wavelength/absorption. The goal was to attack both the root and the leaf. There are more mechanical methods using steam + a spike that work very well too.

Here's a very (terribly) high level technology overview from the company we were working with: https://g-neighbor.com/gni-technology/

> The reason a plant is green is because it reflects green light and for photosynthesis a plant uses blue light. Overloading the blue frequency range disrupts the enzymes in the photosynthetic process, which cuts off the food supply to the plant and it dies. Some herbicides overload the metabolic system of the plant and makes the weed burn from the inside out. I thought that overloading the photosynthetic system would maybe do the same thing.

That’s very cool. Still working on farm space projects?

Not at this time - founded a fintech company instead.

I do miss hardware though!

> Question about this tech is, does it really kill weeds, or just defoliate them?

Question from a non-farmer: if it were cost-effective to defoliate weeds once per day, would that be good enough? I.e., I would think that totally prevents the weeds from thriving.

It depends on the weed. Overall you’re going to put pressure on it. With conventional fields like the ones pictured in this article, there’s often no ground cover. If you have a ground cover (say, clover) and you’re defoliating the weeds all the time, then you have a chance.

But some weeds, like invasive blackberries here in the Pacific Northwest, are extremely resilient and will shrug off defoliation.

Thanks for the info!

> But some weeds, like invasive blackberries here in the Pacific Northwest, are extremely resilient and will shrug off defoliation.

How does that work? I thought all plants (by definition?) need photosynthesis to stay alive in non-dormant states.

Do blackberry plants have some way of getting energy other than photosynthesis in their leaves?

Invasive European blackberries can also draw energy directly from the anger of gardeners vainly attempting to remove them.

If you keep cutting it off and have control over its whole potential growing area it will eventually die. It has some scary properties like roots that travel 20 feet easily and spring up a new independent plant and a insane growth rate once it sprouts. If it’s out of control you probably need someone to first destroy the hedge with a flamethrower

They grow these enormous and very resilient root balls under ground. I was taught that the most efficient way to kill a blackberry bush was to remove the structure above ground, locate a central stalk that leads under ground, and stick the end in kerosene or gasoline I don't remember which. That gets sucked up into the root ball and kills the plant. Never done it myself, mind.

Sure that is a one-stop permanent solution for a plant which grows back its leaves very quickly but the question is, how can the plant thrive enough to harm other crops if you defoliate it every day?

What? I'm telling an anecdote about blackberries over here. It's not commentary on the broader topic.

If the crop is near a bush, then the plant can happily grow in the bush and send roots to the crop.

I’m not a botanist but my understanding is that the root balls hold a ton of energy and the plant can just keep sending out shoots, and spread underground for fairly long distances. So, eventually that energy source will run out, but they’re invasive because of how resilient they are at finding sneaky ways to hide and appear.

> I thought all plants (by definition?) need photosynthesis to stay alive in non-dormant states.

There are parasitic plants that don't photosynthesize at all, like the Monotropoideae and Rafflesiaceae.

Among photosynthesizing plants, rhubarb can be "forced" (grown in complete darkness to reduce bitterness and get an earlier harvest), and potatoes can sprout if left in the pantry for too long.

Mechanical parts have more ways to break. All of those rocks in the field are just waiting to crunch things up.

How long will weeds live if they continuously get their foiliage removed? How about when part of the root (closest to the surface) is damaged/burned as well?

I really don't know, but I imagine that if such a robot went out into the field many or most days, there wouldn't be much of an issue. The weeds would continually get weaker.

Tractors break, sure, but steel implements have been doing a pretty good job of handling rocky soil for a very long time.

Fixing a blue steel corkscrew seems easier than debugging a laser.

I very highly doubt that a farmer is generally going to debug a laser on the spot. That requires a skill set not really needed to be a farmer, and who knows how long it'll take.

Realistically, a farmer would swap out the part.

Any actual work on the laser is more likely to be done by a trained professional, if anything to cut down on accidental burns.

Correct and also you (should) benefit from the main cash crop outcompeting the weed, eventually. Just hope it doesn't go to seed before it dies

> cash crop outcompeting the weed

I thought weed was the cash crop. /s

Had the same thought, also totally agree. Looking forward to seeing more advancements in this field. Pun intended.

Thank you for sharing, unlike people with absolutely zero experience in agriculture speculating about their half baked ideas about it.

Is there a reason you need to mow your slope?

We have an upland prairie which is a vanishing ecosystem which we are conserving. It’s fairly steep. We don’t have sheep but maybe someday we’ll get some. Until then, mowing is a way to keep it healthy and free of invasive plants.

Nice! This is similar to the problem I face with my property, on a much smaller scale. Currently I'm hand weeding the invasive plants, but even for a 100' x 10' patch it's labor intensive.

For areas I manage / garden by hand, I can not say enough good things about this trail tool:


I use it for everything, from digging up plants, to leveraging boulders, to actually maintaining trails.

Is mowing really any less work than pulling weeds by hand? Keep in mind that insects need habitat, too, and tall grasses are key for them. Mowing obliterates insects.

Over time we’re establishing native fescues which don’t get super tall. As well as native wildflowers. All beneficial for the insects. Once established, mowing becomes less frequent. But we’re talking about 20 acre areas, so yea it’s a lot more realistic than weeding by hand.

when i was in HS i worked on one of those crews you see mowing highways and state parks (super good money for a teen but man it was hard work). Mowing on a slope with a large tractor can be very dangerous because you can tip over. Automating away the danger to a human may be worth the effort in its own right.

They make remote control (wired) mowers specifically for this use case. I've also seen some steep-slope brush mowers from Swiss companies, but those are hand-operated. Not surprisingly, the Swiss have a lot of farm equipment designed for use on very steep slopes.

The real problem is tractors aren't designed correctly for slopes, you don't need quite automation as much as a vehicle with lower center of gravity.

Defoliating weeds instead of picking them by the roots is a great way to end up with a field of weeds.

Depends on the weed and the ground cover the weeds are competing against. For many weeds, simply mowing them in the presence of decent ground cover is completely effective, as they’re optimized to grow tall quickly - defoliate them at the right time and they’ve already spent their energy reserves.

This sounds like a much more efficient way to practice monocropping, enhance fragility, and capture economic value by increasing dependence on non-local food sources.

In other news: https://www.sciencemag.org/news/2021/04/pacific-northwest-s-...

The use of this technology is orthogonal to whatever other issues you may have with modern agriculture.

Despite growth in organics, 99% of farmland in the U.S. is still conventional. Anything that reduces chemical herbicide usage on that farmland is a good thing.

The right way to farm is to work with other plants and the weeds themselves to regenerate the soil. The idea that weeds are pests that need to be eliminated at all costs is antithetical to proper regenerative farming. Anything that enables or propagates industrial farming is genuinely a bad thing that we should not support as technologists, it is not at all orthogonal! "A little bit less of a bad thing" still leaves us in a bad place.

Yup, agreed. Speaking to others listening in: some weeds are edible, and other functions as part of ecological succession. We could be using farming practices that goes with that, rather than against it. Instead, industrial farming is optimized to produce single crops conforming to consumer expectations, cost-efficient harvesting, and durability for long transports and storage. Industrial farming is not optimized for nutritional value, freshness, and resilience against environmental stressors. The system by which 8 billion people are fed resembles a Ponzi scheme in which we are borrowing against future generations ... and with the increased variability in weather and water rights wars, it is only going to get more, not less, fragile.

Example is the dandelion. Besides having culinary and medicinal value, it acts as a pioneer species for depleted soil. Killing it with more roundup or zapping it with a laser, and then contaminating the land with more fertilizer will just encourage more dandelion growth. The land and ecosystem is signaling a fertility issue, and our present practices work against it.

I say this even though the common mallow is the bane of my existence here in the lower Sanoren ;-)

Why did you link to an article about pacific northwest indigenous forest gardens.

It seems like you are making an implied argument. Not stating your argument ensures it can't attacked. Not wanting your argument attacked suggests it is weak and won't stand up to scrutiny.

You caught me red handed. I have made explicit arguments elsewhere on HN before. It usually involves writing a lot of exposition, starting with identifying and deconstruction of the default paradigm that some people call “Value Extraction”. It can get very in involved.

This time, I was adding a subtle sarcasm, which I knew a small minority of HN readers will pick up. They are already familiar with food forests, restoration agriculture, and some may even be familiar with regenerative paradigms.

But let me see if I can condense this into something explicit:

Rather than eliminating everything but the monocrop, there are alternative forms of agriculture that takes advantage of the synergies that come from companion planting, “guilds”. Specific combination of plants planted together can be put together to reduce ecological invasions, or create an ecosystem that can produce harvestable food items for most of the year. Some combinations can mutually resist pests and diseases. Others can be combined to take advantage of vertical spaces (canopy layers), or to effectively modify the local hardiness and heat zones, or modify wind conditions.

Designing such a thing can get very complex. Some designs, however, become so resilient, they continue producing human-harvestable food despite being abandoned for over 150 years.

The Pacific Northwest is not the only region where indigenous people have used these practices. There is quite a bit of anthropological evidence that this is a wide spread practice ... what is novel with that article is that ecologists are acknowledging that this had happened in the Pacific NW.

>Rather than eliminating everything but the monocrop

No one is suggesting this be done. Fanatical monocropists don't exist.

Your argument is based on a false dichotomy. We can make monocrop farming more efficient and environmentally friendly (eliminate pesticides) while also exploring other systems.

> We can make monocrop farming more efficient and environmentally friendly (eliminate pesticides)

More efficient and environmentally friendly are contradictory in terms of monocrop

> Fanatical monocropists don't exist.

Uh what? What about the almost 100 million acres of corn in the US?

I am not sure how you define monocropping, and so I will write what I think it is.

By “monocropping”, I am talking about planting a single cash crop on a plot of land. A farm may operate multiple crops, each on its own field. They might rotate crops with cover. Some might even interplant crops in rows, or even between plants (as cover). Each field is intended to grow and harvest a single cash crop. You don’t even want to different varieties of the same species close together because you want to ensure a consistent produce. Many restaurants demand consistency (easier to consistently produce dishes when ingredients are of consistent size and quality), and consumers have been conditioned to take a few varieties as representative. (Example: many consumers think of Roma tomatoes when they think of “tomato”).

The combines are designed to till, sow, and harvest in generally linear rows, for a single type of crop. The field is plowed in a way to be level, so that watering can be as uniform as possible. Soil amendments, such as fertilizer, is relatively simple to apply.

So as an example, an old Native American practice called the Three Sisters: corn, squash, and beans, can be planted together. These three crops, when planted together, includes a plant that will fix nitrogen (beans), as well as a ground cover (squash) that help retain moisture. The corn stalk functions as a trellise for the beans, and help shade it. Beans can be harvested early in the season, corn and squash later in the season, with squashes working well as a dense store of nutritional energy that will last through the winter. It also happens that there are lots of culinary dishes that also combines these three plants.

I know of no industrial farming practices or equipment that practices the Three Sister. Driving a harvesting tractor through there will destroy the squash, and probably the beans too.

There are other combinations such as, tomatoes and basil; strawberries and asparagus.

When I use companion planting in my garden, I noticed a few things. First, harvesting is more difficult. You have to be able to identify each plant. I had to look through the different canopy layers to see what is going on. But it was also my first attempt, and I think I can design it a lot better.

When done this way, yield are not as good for any single crop (if they don’t synergize on nitrogen fixing). But I also wonder if the total crop output across all harvestable plants are better.

That is just for annual crops. There is a similar practice that can done for perennial crops. This is what those “food gardens” mentioned in that article, though they are usually called “food forests” by enthusists.

One of the things about this method of farming is that it is intended to decentralize food production, letting the bulk of the food production be local, in order to reduce transportation. You can then grow varieties that taste better and are more nutritionally dense, instead of being bred for staying edible and transported by refrigeration. (I am looking at you, roma tomatoes and hothouse tomatoes).

During the years of international sanction, Cuba had to develop self-sufficiency with use of very little fertilizer (no oil imports), no gasoline driven farm equipment (no oil), and as little motorized transportation as possible. They used a variety of methods, some drawn from modern science and some drawn from older farming practices; but whatever it was, they could not simply use pesticides, fertilizers, or refrigeration.



When you said "Rather than eliminating everything but the monocrop" I assumed you meant society eliminating non-monocropping farming methods.

Is that what you mean or did you mean something else, like the machine could eliminate some non-monocrop plants but not others?

Ah I see, I did not write that well. Let me see if I can rephrase that: in the practice of monocropping, we eliminate everything but the cash crop in order to optimize yield and profit. Everything that is not the cash crop is a weed.

Rather than doing that, we can use companion planting, integrated pest management (chickens, ducks, etc) on smaller fields.

Smaller fields and multi-crops don’t have the same scale as the current farming operations. But the idea here isn’t to replace a large commercial monocropping farm with another large commercial polycropping farm. Instead, we are replacing large monocropping farms with many smaller farms using these methods, and locating them to population centers.

As far as the point you brought up about society eliminating non-monocropping methods ... I don’t think there is a need to have our society use laws to force monocroppers out of business. Instead, I think it is better as a grassroot effort with people practicing this, as a sort of ecological succession —- people starting up home gardens and neighorhood farms to supplement existing food supplies, with monocropping dying on its own.

It does not even have to happen in affluent places first. There are many urban food deserts where people can only get expensive, nutritionally-poor food from convenience stores. You get some small operations there, even hybrid methods, with community involvement and it can change a neighborhood. (Examples: Urban Farming Guys out in Kansas City, and Brad Lancaster out in Tuscon, AZ)

There are plenty of commercial farmers who want to get out of the debt trap they got into with cash cropping. Some even tried their hand at organic farming ... but some will bring their big farm monocropping mindset into it and fail. I think many farmers (and not the agricorps) love working the land and feeding their society. The ones that don’t have been leaving farming for easier work.

But I do think there are things our society does to keep agricorps going. If food decentralization reaches a certain point, I think those agricorps will fight back to retain control of their markets.

And to address the several commentors who say this technology is orthogonal to agricultural practices:

An alternative use of ML, CV, and robotics, might be something that can observe and identify all the participants in an ecosystem, such as food forest, and use ML to sketch out the possibility space in which viable cross species synergies can enhance an ecosystem. This could be used as an aid in designing a food forest.

One use of ML, accelerates ecological degeneration. A different use of ML accelerates ecological regeneration. I don’t think the latter involves a weed killing robot.

I agree, ML is a tool that can be wielded in many ways. Another ML use would be to find new combinations of crops that work well together. We can have a "genetic algorithm" design gardens, with successful ideas being replicated in future gardens.

I'm so tired of these endless orthogonal attacks on technology that is aiming to help the environment. Of course things are perfect. This could reduce the usage of tons of dangerous herbicide.

Go start your own thing instead of lazily smashing others.

Elsewhere in this thread someone suggested the opposite. Presumably you could attempt to implement this either as "destroy all species other than C" or as "destroy only species W". I don't know how to think about which, if either, this approach is more suited to.

That was my thought as well. Thanks for sharing the interesting article

Yeah, this robot is an example of tech being utilized to extract value. The ML, CV, and robotic tech tickles my inner geek, but tech was not designed or deployed in a regenerative way. There are many things we can do to have more resilient, restorative, regenerative agriculture ... that often don't involve much high tech at all.

So what is a more worthwhile use of high tech? Being someone in tech and having greatly economically benefited from it, I've recently been reframing what I know about software tech, startups, innovation with a regenerative paradigm. I am still making my way through Carol Sanford's work on regenerative paradigms and figuring out a lot of this stuff out. So far:


Christopher Alexander had introduced the idea of creating pattern languages for people living, working, and playing within building architecture so that they can modify and design their own living spaces. What we ended up with are cookie-cutter housing in suburbia. His work greatly influenced OOP and Human-Computer-Interface design (see his 1996 OOPSLA Keynote) ... and what we ended up with is the Gang of Four, Apple products, growth hacks, and "user engagement".

What we don't have are individuals, families, and communities having computer tech that can customized by the users for what works in their local environment. Smalltalk was designed with that intent in mind, but our legacy from Smalltalk is the Gang of Four and OOP "design patterns".


There was a recent article posted here. It was a fictional interview, the premise being someone from a parallel world where software design was elevated on par with science and art, and not merely engineering. They start with a "design brief" rather than engineering requirements. I lost track of that article.

Just some evolving thoughts.

> Yeah, this robot is an example of tech being utilized to extract value. The ML, CV, and robotic tech tickles my inner geek, but tech was not designed or deployed in a regenerative way. There are many things we can do to have more resilient, restorative, regenerative agriculture ... that often don't involve much high tech at all.

I have this struggle often. Before software engineering, I was working on permaculture farms and apprenticed as a natural builder (cob mostly).

Have you looked at sandstorm.io ? I have been looking at it as one of the pieces for “regenerative software”.

I still have no idea how to make a living from this though.

It's all fun and games until the selection pressure evolves weeds that say "human" on their leaves to fool the AI.

Nah, they would be cultivated and breed for other texts or be driven to extinction like those trees which evolved numbers https://wiki.lspace.org/mediawiki/Counting_pines

Farming goes from broad selection against pests with chemicals to introducing a massive selection pressure towards biological mimicry and adversarial image attacks.

I wouldn't worry about it too much, for a couple of reasons.

First, the evolution of the computational model is simply faster than biological evolution. The computation model is going to be inside biological evolution's (metaphorical) OODA loop. Humans are going to be helping, too, it's not like it's just going to be up to the deep learning algorithms on their own.

Second, for most weed plants, they aren't just a couple of genes away from mimicking corn... they're probably dozens or hundreds of genes away from mimicking corn. Evolution is OK at adapting current things to new uses, or doing a massively-parallel search on what you can do with just a tweak to a gene, but if the task can't be done with one of those things, it just loses and the organism dies. Or, to put it another way, it's good at climbing slopes one step at a time, but if you present it with a cliff it just fails.

It's essentially the same reason why nothing has evolved a resistance to a human gardener yanking them physically out of the ground and leaving them to die on the concrete... it's not just a matter of tweaking a couple of genes for that. This robot presents an exceedingly harsh selection landscape for a weed.

> Evolution is OK at adapting current things to new uses, or doing a massively-parallel search on what you can do with just a tweak to a gene, but if the task can't be done with one of those things, it just loses and the organism dies. Or, to put it another way, it's good at climbing slopes one step at a time, but if you present it with a cliff it just fails.

Not the case. https://en.wikipedia.org/wiki/E._coli_long-term_evolution_ex... gives us a concrete example of this; the ability to metabolize citrate required two mutations, the first of which isn't useful on its own.

As long as a mutation isn't drastically harmful, it may persist in the population to be built upon later.

> It's essentially the same reason why nothing has evolved a resistance to a human gardener yanking them physically out of the ground and leaving them to die on the concrete...

They have evolved quite a few of these. Try and clear a lawn of dandelions and you'll discover they regenerate from their taproot if you leave any of it in. https://en.wikipedia.org/wiki/Vavilovian_mimicry indicates another, where false flax has evolved based on a human invention, winnowing machines, to evade them.

Rye grain evolved so effectively based on human activity that it went from weed to useful crop, even.

> Evolution is OK at adapting current things to new uses, or doing a massively parallel search on what you can do with just a tweak to a gene

I think evolution over longer timespans is the algorithm that keeps giving. In a single run it created all life and current technology. Evolution is radically open-ended, and that's how it gets around deceptive search spaces.

And eventually that mimicry just creates us new secondary crops like oats: https://en.wikipedia.org/wiki/Vavilovian_mimicry

You're missing another remarkable ability of nature - coevolved species with humans. In essence, they keep evolving until humans find a use for them in their lives (be it personal, medical, industrial, etc) and then end up caring/nurturing them. Or another way to phrase it: The weeds evolve so they don't compete with the crops the human grows.

A few examples come to mind:

- "Clovers" fix nitrogen in the ground, other plants can take that nitrogen from the ground, some gardeners now intentionally keep this weed to benefit the plants they are growing for crops. I think the term is called "rotating cover crop".

- "Comfrey" has incredibly long tap root which mines minerals and stores those in the leaves. Gardners/farms can plant comfrey, chop the entire top off, the decaying leaves release the minerals into the surface soil for the surround crop plants to utilize. Without comfrey, those minerals would of been locked away from the main crop. After the comfrey is cut, it grows back the leaves again and the gardener/farmer can repeat this cycle.

- Some weeds attract beneficial insects to the garden/farm and thus benefit the main crop. Wise gardener/farmer would keep these around. For example, more bees means more insect pollinated fruit to be grown.

The pattern with highly competitive specifies results in either extinction of one of the species or a mutual beneficial evolution. For example:

- "Bull horn Acacia tree" - Ants and the tree have co-evolved that in the present they are highly depend on each other for survival. In the history, at one point when the ants and tree were introduced the ants brought a lot of acacia tree destruction followed by ant death because of loss of food source - but over time the trees that benefited the ants got selected along with the ants that were compatible with the trees.

If you'd like to learn more on the topic of coevolved species, I recommend this 30 min video on the topic https://youtu.be/hCAvBmY7ZgA

I don't know what this robot will do, the story is being currently written so we will see what happens!

I stopped reading at "Farming robot kills 100k" and ran to my prepper basement.

Or evolves the weeds to grow their own lasers.

Just a shinning/mirrory-silvery leaf.

Or until gamma rays flip a bit and the code becomes `kill !weeds`

If that were possible, wouldn't it have already evolved that millenia ago in order to evade humans pulling it up by the roots?

They absolutely have, in various fashions. Try getting dandelions out of a lawn - miss a bit of the taproot and it’ll come right back.

Elsewhere on this thread is a great example of this in practice. https://en.wikipedia.org/wiki/Vavilovian_mimicry

> Seeds that are thrown the same distance as flax seeds have thus been selected for, making it near impossible to separate the seeds of these two species.

Are you thinking of the CLIP adversarial attack where they wrote "iPhone" on a potato?

I know this is supposed to be a joke, but come on. In what possible world would this ever happen? They're not dragging this over people lying down in a field and I doubt the technology is even made to not target humans because they're never in target range.

True, the weeds will evolve to have the text "Corn" on them.

And eventually the humans will too.


Typographic ML attacks work and are pretty funny.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact