Hacker News Reader

Comment by dicroce | Hacker News Reader

Many crops would not need so many pesticides if they were grown in close proximity to the other plants they naturally occur with... in some cases this is because other plants attract predator insects that keep other insects at bay... humans could never farm this way because harvesters just chop everything down at once...

But a robot capable of recognizing all the plants could tell... and could selectively harvest just what is needed.

I suspect this is a very difficult problem however just because of how varied individual plants can be... telling plants apart can be hard even for trained humans.

View on HN

I suspect this is a very difficult problem however just because of how varied individual plants can be... telling plants apart can be hard even for trained humans.

Actually, it's not. Despite all the hype surrounding deep learning, tasks like telling similar looking plants apart is precisely what's it's good at. In fact, it will probably be better than most humans at this task. Moreover, it's simple to implement. You take off the shelf model (e.g. ResNet-50) trained on ImageNet, and finetune it on a dataset of specific plants you want it to recognize (~1000 examples of each plant). Then you can run that model on an iPhone (or even on much cheaper hardware)

A much harder problem is to make the robot move through the terrain without getting stuck. But even there, looking at what Boston Dynamics is doing, it seems like we are not that far off.

Robots with legs are probably not an option, they are both costly in production and energy usage. Something more akin to https://farm.bot but on a larger scale would probably work better. Maybe a hybrid solution like a CNC bot built on top a frame with wheels that could simply drive around a field and work on different quadrants would be a way forward.

I am currently making a suspended robot to work on that exact problem. I hope to achieve what Farm Bot does but on a much higher scale. I live in Japan where most of the fields are ricefields. Wheels or legs are not really an option under these extremely muddy conditions.

Would love to hear more about this

I wish this post happened a week later! We are currently making a website, it should have been online last week but we missed the deadline. I have a small working prototype (2x2m area), I am currently changing some parts so that it can be made 100% with off-the-shelf parts. Next step is trying it on a 10x10m area and then see how far it can go.

Without making it particularly sturdy, it was able to lift 4 kg and has a pretty good movement repeatability even if totally "blind" (I am currently working on its vision system), two things that drone would have a hard time duplicating in outdoor conditions.

I can PM you later when we have all the things set up if you want. Or we can just chat here.

I'm building something similar but for litter collection.

In order to make the system mobile I'm making a sort of tiny tractor that relocates the bases of the support struts (I use a tripod.) You can have separate symbiotic fleets of static "heads" (tripod+wirebot) that are moved by mobile "rovers".

This lets you optimize the system a little better, since most of the time the head units will be working and largely stationary, so a small fleet of rovers could be enough to handle a large fleet of processors.

For ricefields I think you might sink stepping stones in a triangle "grid", maybe concrete cones about a meter long and 10cm at the base, inverted and stuck in the mud. The distance between them would be fixed by the span of the suspended robot.

You would also want "vascular" conduits to transport material to and from the working heads. These aren't necessarily physical tubes or conveyor belts, they could be made of simple cargo drones like streams of ants.

You could make bridge or rail units that fit between the stepping stones for temporary reconfigurable transport networks. (If your cargo drones are sure-footed and reliable these could be simple planks.)

Also, have you heard of Hangprinter? https://vitana.se/opr3d/tbear/ (I have no affiliation w/ it.) It's a "RepRap hanging from the ceiling"

Interesting! Actually my dream would be to make a system deployable by a fleet of drones. Rovers are actually complex to move, especially in an uncontrolled environment with lots of plants.

I don't think any rice farmer would like to put concrete blocks in their fields unless every task is reliably automated: they depend on machines to prepare, plant and harvest the rice and a stone in the middle of a ricefield would make these steps harder.

I know that there are several projects using wirebots for 3d printing. It usually allows for a bigger printing volume than regular CNC-like designs, and I think they must be able to have a high precision too. I am wondering what is the limitation there.

I had a similar idea of making a suspended robot farming robot (e.g. it may hang in between two fences of a backyard or between trees...)

Do you have any contact info?

Sure: info@fae-bot.org (the website is mostly a placeholder right now with old pictures of an early prototype)

Interesting, so how does the winch system work is it actually pulled and released from each pole base? Are they going to be modularized so they can pass tools along multiple such systems? Have you thought about harvesting and collecting?

I use a winch based on a stepper motor. Everything is modular so that it is easy to make a robot with just 3 wires or with 8 for 6 DOF experiments.

Passing tools among a neighboring system is going to be hard as there is a "dead zone" bordering the rectangle drawn by the pole but there may be a way to smartly cross the wires so that two system may share an area where they could pass around tools.

Harvesting is actually, counter-intuitively, a low priority task IMO. This is usually a single day of work, and this is the one with the highest stake, where things going wrong can jeopardize your whole harvest. People will want to be present for it I think.

What needs to be automated first are the tedious tasks that need to be done repeatedly : weeding, monitoring, watering.

About harvesting, a farmer friend proposed something interesting: just pass bags/baskets from the field's side to the place when the human is inside the field. Ricefields are very muddy and if you work in the middle of one, doing back and forth to the border is really tedious.

Congratulations! Please use the email address in my profile and feel free to put me on any email list you have.

hey I'm also interested in this. Mind shooting me an email when the website is ready? hello [at] myusername.com

See my comment above, the original steam trackers were too heavy to put on a field, so they used a system of wires to pull machinery around. Two trackers drive in tandem down the sides of a field, with a wire between them. The machinery suspended from the wire.

You could put it on a radial track built like center pivot irrigation systems. Maybe even on the same construction if you already have that.

This is a billion dollar business, if it can be made to ride on existing systems.

Imagine being able to plant, weed, and harvest strawberries from a center pivot. Huge.

I'm betting you'd still need to install a new track designed for the purpose, but you might be able to clamp it on to the existing framework. Irrigation stuff is below and would be tough to move around, maybe the robots dangle off the side?

Legged are impractical now and for the near future, but I suspect the cost will decline substantially within 10 years.

Why does it help to be trained on imagenet first? Also are there any tutorials to teach you how to do what you’re describing?

It’s called transfer learning. The base layers of a model trained on imagenet are general enough that they’re helpful for the specific task of processing plants. The benefit is imagenet has much more training data which is why you’d train on that first and then fine-tune the model with your specific data. Fast.ai is a great beginner friendly resource.

It’s a hard problem I know BASF has invested a lot of money working on this very problem.

> humans could never farm this way

... at scale. Humans do farm this way -- I've seen fancy shade-grown coffee being grown in what looks like a jungle. But it requires a huge amount of manual labor, right now.

I think we'll get there at making machines do this. Deciding whether a leaf is one of the 4 crops you're growing, or not, sounds much easier than distinguishing all the close relatives of wild plants.

> Humans do farm this way

Yes.

There are two different senses of "efficiency" in farming (or other production). One is the amount of food you get per unit of human labor input. Supporting a fully urbanized population requires this metric to be high.

But the other sense is the amount of food you get per unit of land input. A society that maximizes this metric will produce more food and have a higher population than a society maximizing the first one. But most of those people will be subsistence farmers who don't earn much more than the marginal extra food they provide by intensively caring for the land. This is why historically most of the population -- anywhere -- was rural peasants.

The grotesquely inefficient capitalist who uses a workforce of 10 people and 300 acres of land to produce the yield of 50 acres of land is part of a "wealthy" society -- everyone else has more food because of his very low need for labor. But it's also a very low-population society. In a more traditional society, all the land would be farmed efficiently and the population would be much larger, but non-farmers would have much less surplus to capture.

There are a lot of choices which go into all of this, too. One interesting story is about which poor countries tried to copy 20th-C US farming, and which did almost the opposite.

The first path was a mistake -- it looked modern, all those big machines, but was optimised for near-infinite land and very expensive labor. What poor countries typically had was the opposite: plenty of peasants and no room, and no diesel. So the winning strategy was actually closer to a million vegetable patches, with incentives, training, and transport to get this to the cities & export. An acre of vegetable patches can easily produce 100x the market value of an acre of wheat. Getting this right was a big part of how Taiwan & Korea for instance kick-started their growth. Getting it wrong was part of what killed the USSR -- if I remember right, they never even equalled the 1917 harvest.

interesting anecdote. any recommended sources?

Seeing Like A State goes into a fair amount of detail.

From my notes:

* Soviet collectivization. Inspired by huge mechanized farms in the US. Vision of doing for farming what assembly lines did for manufacturing. Both Soviet and US attempts to do so failed badly. Party resorted instead to martial law and grain seizures, creating peasant uprisings. Forcibly relocated peasants into standardized, pre-planned mega-farms which were ruled by agricultural specialists. Ignored local conditions, vastly over-simplified and over-abstracted different areas of land and differently-skilled populations. Pressure from above to deny failures led to plans quickly losing contact with reality. Peasants gained most of their food by farming their own private plots in their little free time. Peasants effectively became indentured slaves, stripped of any cultural institutions that might be a focal point for rebellion, and naturally responded with terrible productivity. More deaths from starvation that WW1 and civil war combined.

* Compulsory villagization in Tanzania. Similar story. Government (with support of the Western world) wanted to modernize the peasantry. Forcibly relocated peasants to standardized, pre-planned villages. (Was supposed to be voluntary, but top-down pressure for results led to initiative-taking). Peasants were moved vast distances, rendering their deep local knowledge worthless. (Peasants don’t come in standardized, fungible units). Forced to apply Western farming techniques which failed badly in the local climate and ecology. (The peasant practices of dense polycropping turn out to be much more effective in climates with high primary productivity than monocropping and ridging as is common in the West). Authorities heard reports of atrocities but insisted they were isolated cases.

* And again in Ethiopia. Lead to widespread famine.

"You should use wheat-growing techniques that do work for us, but don't work for you" isn't quite the same message as "you should grow wheat (using techniques that work equally well for us and for you) even though wheat is only economical for us because we have a lot of land and no people to farm it, whereas you have a lot of people but not so much land".

There are varying levels of communist madness, from "you must pursue this strategy, sub-optimal for conditions here" all the way to "you will stay in this barbed-wire pen until you all starve". In the 60s/70s Tanzania was further along this line than was Russia.

For downvoters (who perhaps don't like my swipe at commies): my point here is that not all of these were honest if misguided attempts to improve farm productivity.

Some, like in Ethiopia, were closer to ethnic cleansing -- deliberately moving people (esp from Tigray IIRC) who were violently unhappy with the government, to less fertile places where they didn't understand how to farm, especially with new crops to boot. But by claiming this was a modernisation scheme was excellent PR, people like those, and the resulting famine was seen primarily as a great tragedy, and brought in lots of aid money.

How Asia works by studwell. It’s also a Bill Gates book recommendation.

This is a great book, and easy to read.

For russia I also like Yegor Gaidar, there's a book but the summary is http://www.aei.org/feature/the-soviet-collapse/

I wonder if the Communists just got bad timing. I am not saying Capitalism isn't better - but maybe they could have made it if they'd had robotics.

Communism and capitalism will hopefully be made irrelevant by advanced robotics. A post-scarcity economics does not care much about structure of ownership of the means of production.

Regrettably, there is no such thing as post-scarcity society. The exponential nature of biological reproduction ensures that no matter how much carrying capacity one builds, it will be filled to the brim in a geological blink of an eye.

Hopefully a society that is liberated from the burden of work will devote much more time at understanding and solving its political problems and should see unchecked growth as an evil to fight. Especially if we also get very extended life expectancy.

There have been plenty of societies historically in which very little work was required in order to live normally.

What people did with the resulting free time was generally to make war on those nearby.

That's simply not true. Until recently agriculture was extremely labor intensive and only an elite caste could have free time.

You're thinking of agriculture north of the tropics.

But there have been many societies which

- didn't engage in agriculture at all, or

- engaged in tropical horticulture, which is low-effort

And those societies (plains Indians / Australian aborigines / sub-Saharan Bantu (they're farmers! But they're not labor-intensive farmers) / central Asian Turks and Mongols / etc. etc. etc...) are characterized by near-constant warfare.

Not all societies are ancestral to yours. Most aren't.

Do you have any source to your claim that these societies required little work?

And seriously, claiming that a society will engage in warfare by comparing it to an example that is as remote to it as possible in terms of technological development and culture (we do view warfare as negative, not all societies do and ours used to not do so) is not exactly convincing.

That pre-colonial farming in Africa required relatively few hours of labor is well-known. I don't have a great link but "female farming system" is a term to start from (seriously, I didn't make it up).

The reason, if I understand right, was that the limit on human population was set by disease not labor. In northern climates (where fewer of the diseases we evolved with thrive) instead the marginal farmer was on some stony hillside from which maximum effort could only just produce enough calories for winter.

I'm less sure whether such societies were involved in more warfare, it's possible (the men had time on their hands & soccer hadn't been invented...) but I don't know the data.

I also don't know how to extrapolate this to the future of our society.

> There are two different senses of "efficiency" in farming (or other production). One is the amount of food you get per unit of human labor input. Supporting a fully urbanized population requires this metric to be high.

> But the other sense is the amount of food you get per unit of land input. A society that maximizes this metric will produce more food and have a higher population than a society maximizing the first one.

Are you suggesting that the reason the US/Europe is not at the Malthusian limit / has a low population growth is because we don't grow enough food? You'll have to pardon my incredulity.

Your point is strictly true - we are growing less food than we theoretically could grow, but that seems beside the point: we still have an enormous surplus of food.

> Are you suggesting that the reason the US/Europe is not at the Malthusian limit / has a low population growth is because we don't grow enough food?

Population growth isn't being restricted by food availability. As to the other questions, I wouldn't want to argue causation in either direction. I'm pretty sure it goes in both.

If we started to reproduce at the level supported by our food production, we'd end up in the Malthusian equilibrium.

The west has been decoupled from Malthusian equilibrium for two to three centuries.

At earlier times, say 800 years ago, I think it can be argued that population densities in Europe were lower than in China because farming was optimized more for labor than for land -- plowing with horses is less work, but less food, per area, than planting rice by hand. But it's super-hard to figure out cause & effect obviously.

> but non-farmers would have much less surplus to capture

As someone who is grateful for computers, toilets, psychotherapy, punctual trains, and an efficient medical system... I am very grateful for our ability to produce food without high labor inputs.

Don't worry, it'll be substituted with agricultural robots soon enough and we can have the best of both worlds.

A lot of the permaculture/edible forest movements are dependent on cheap labor, usually of students coming to learn the techniques or volunteers into a more sustainable lifestyle.

I live close to both a rural hackerspace and a permaculture community. I am currently working on a suspended wirebot to help them automate the most tedious tasks. Pest removal, weeding, monitoring seem to be the low hanging fruits there.

> But a robot capable of recognizing all the plants could tell... and could selectively harvest just what is needed.

You don't even need a robot for this. The city of Edmonton, Alberta has deployed a fleet of small weeding machines to curb pesticide use -- to wit, specially-trained weed-eating goats.

Could you provide a link please? On pasture land I see the opposite problem - the animals eat the 'best' stuff (clover, grass) and leave the weeds (thistles, nettles, buttercups, ragwort etc). (Although partly this is because on pasture land, weeds are 'what animals do not eat' :)

https://www.edmonton.ca/activities_parks_recreation/parks_ri...

Goats are known for their ability to consume a wide variety of plants, especially weeds and shrubs that other animals won't touch. These goats have been specially trained to target noxious weeds and are constantly watched by humans and dogs.

Unlike human, computer is exceptionally good at remembering stuff. No need for plant recognition when you can brute force remember where you planted each individual seed.

>Many crops would not need so many pesticides if they were grown in close proximity to the other plants they naturally occur with.

This is the strategy JM Fortier employs in his gardening technique. In my retirement I plan on attempting to implement the market garden (a smaller version).

http://www.themarketgardener.com/

>> humans could never farm this way because harvesters just chop everything down at once

Really? You think a mechanical harvester cannot be built that could handle two types of plants at once? Forget deep learning and imaging technology. Plant alternating rows and build a harvester to accommodate, straddle, one while working the other. No fancy vision tech needed.

It's certainly possible to build such a machine. But it would be more mechanically complex and expensive to build and operate than existing harvesters. And it would be locked into two specific crops which might not be the combination that farmers want next year. It's hard to see how machinery like that could ever be economically viable.

Possible? Try a standard farming technique called "intercropping". Often a single mechanical combine harvests both plants at once, with the crops separated from each other mechanically later. Other times the plants can be separated by height and harvested on subsequent passes.

https://en.wikipedia.org/wiki/Intercropping

https://www.grainews.ca/2013/10/08/six-things-to-consider-wh...

A robot with computer vision could easily tell apart maize from soy from squash (different colors, react under polarized light differently, leaf size, plant structure, etc). The real problem is that 1. squash has limited commercial value. 2. corn grows to 8' tall while soy is only a couple feet. The corn would block all the sun from the soy. When you plan a field full of corn it all grows at a uniform rate so each stalk gets an even amount of sun, rain. If you want the advantage of both crops, you rotate every year or two. That way the soy can perform nitrogen fixation with various bacteria and produced essentially free nitrogen fertilizer for your corn.

I suspect that a robot capable of doing that would nearly as expensive as a human doing it.

The first one would be far more expensive than a human. The 100th one would be far cheaper.

More like the 100th. generation ...

It would also be incredibly slow. You might cover a few acres per hour and would need dozens of these operating in tandum to cover a large farm.

Or a robot simply removes the pests one by one. If you’ve got enough robots it shouldn’t be a problem.

Once you're making your robots fight insect/animal pests, it's probably more efficient to attract natural predators. They're like highly-optimized, self-maintaining, self-constructing robots already.

This is the awful irony of a lot of these farm robots. Like the pollination robots, for example. Maybe I'm just being shortsighted, but it seems like we already have pollination robots, called bees. Native bees don't even require a hive or beekeeper.

But, apparently, it would appear "not using neonicotinoids" is a lot harder than "designing & manufacturing millions & millions of tiny robot bees"

Bees are not always efficient pollinators and they can also be most costly.

More costly than building thousands of tiny robots or pollinating plants by hand?

You need to consider the costs of all these robots. Farmers would in razor thin margins.

how fast and of what size must such a robot be so that it becomes cost efficient?

There are existing prototypes of laser systems for killing fast moving insects such as mosquitos. Imagine a device on a fence pole covering the surrounding acre.

There are plenty of slow moving pests, weeds and the like as well.

Or we could eat less fresh fruit, buy more fruit for baking and already canned fruit.

Where do you think that canned and baking fruit comes from? It's grown just the same.

Presumably, canned fruit could involve less organic waste as the canner can use ugly fruit that consumers may reject for long enough that it spoils.

I don't know enough to guess whether this is more energy 'efficient', given that we need glass or metal containers, in this case.

Yes, but there's probably less loss & spoilage in the supply chain for canned/frozen/dried/etc.

Same pest/weed control, though, and that's what we were talking about.

How about putting different plants in different rows, e.g. plant A in even rows, plant B in odd rows.

If you're cash crop is plant A, then plant B needs to be cheaper than spraying herbicides and the cash value of all the plant A you could have grown instead.

Unless there are laws against spraying those herbicides.

And laws could be drafted once it becomes known that these types of method are feasible.

https://en.wikipedia.org/wiki/No-till_farming

I believe this is called intercropping, although both plants A and B are meant to be harvested.

https://en.m.wikipedia.org/wiki/Intercropping

I definitely agree with your point, there may not be a need for a high tech solution to that problem.

I would be curious what ratio of producing vs protecting rows a farmer would need. Perhaps it’s less than alternating rows, not harming yield as drastically. Perhaps 1/4 or 1/5 would do.

I need a club to talk about this asap

I am currently building a suspended wirebot, I have a prototype that is starting to run nicely, hopefully the website is coming up soon. I'd love to talk about farming automation experiments (that's actually the acronym of our project: FAE), the needs, the dead ends, etc...

Wirebot are cool but invasive (no offense) IMO. I'd be happy with a bunch of wood based rovers. Some tiny helpers to ease and accelerate grabbing fruits/plants. Things super simple but that could turn farming into painless cute thing (if possible, I'm mostly daydreaming).

That's an interesting perspective. Why do you feel them invasive? I have the opposite perception as rovers have the risk of running over valuable plants. Maybe my perception is skewed by the fact that I am surrounded by ricefields and that wheels big enough to work in that muddy soil will basically destroy the field...

Maybe I have the wrong picture but I thought wirebots required 4 poles and diagonal wires, which would clutter the view and feel like band aids over the field (no offense, again, I do not know how it is in reality).

Yeah I didn't mean big machines, don't laugh but I have wood+lego (small, simple blocks, cute) in mind when I think about automating farming with devices. Slightly larger, not toy size, but nothing like the usual farming scale utilities.

Hard to communicate this over text (especially since i'm only imagining). I'll draw some stuff one day.

Reply on news.ycombinator.com