Physical Intelligence's first generalist policy AI can finally do your laundry

Where did the economies get a boost when women joined the labor force? In Germany the post-war economic miracle happened mostly without them.

In the beginning of the 1970s, when women did join the labor force in greater numbers, admittedly the macroeconomic conditions were bad (oil crisis), so it is hard to filter that out.

But still, mostly we have more workers, which lowers wages and leads to the creation of more bullshit jobs. To be clear, also men create and perform bullshit jobs!

Now it takes two salaries to finance a house and a family. Great progress.

At the supermarket, I boycott automated self-checkouts even if the lines are long so the nice cashiers keep their jobs.

The underlying real goal is to have sex with the robots, just as the internet was going to be the gateway to all known information but instead is for porn.

>3. In the western world companies over past fifty years got a free boost as women joined the labour force, and essentially companies were paying one guy the cost of a household now they pay two people the cost of household getting twice the workers for same price.

If demand for work is fixed, then doubling the supply of work (male + female from the same household) would in principle decrease wages (not necessarily by half). However, the supply need not be fixed. This is what I like to call the "musical chairs theory of employment". Let's say there are 2x more working age people in 2024 than in 1954, does it mean each person earns less? When immigrants come to a country, does it mean that each immigrant causes a native to become unemployed?

The evidence tells us that's not true and that there are other changes in the economy that also increase the demand for work. Check the literature of immigration effects (even major short term shocks) on unemployment and wages to see a pretty clear picture. I'm not aware of a study of what more female employment has done (my guess is that it's much more difficult to study, because it's a phenomenon that has slowly happened over decades. As an aside, this has happened in plenty of places around the world, no need to restrict yourself to the western world. Plenty of other places are relevant)

These Physical Intelligence guys are very talented, but they will never raise enough money to achieve their goals. Their problem is saying it costs $30m to develop such a technology, when it will cost at least $1b, and probably closer to $5b.

I recall reading many comments on HN confidently predicting that the moment a self-driving car caused an accident that killed someone, that would be the end of self-driving cars. But while they have caused accidents, and there has been resulting lawsuits and investigations by regulators, it hasn't put an end to them. And with the incoming US administration, I'm expecting far less legal and regulatory barriers to greater use of robots and automation.

Even if they need supervision, watching a robot hang up the laundry to dry is better than having to do it yourself.

Like that old quote: "I love work, I can watch it all day."

How about at first if we put camera on it and then someone from overseas checks around before performoning the task? for cheap but it can make it secure.

I agree. So far I haven't seen production-ready robots doing even relatively simple agricultural tasks, such as picking tomatoes in a greenhouse and taking care of the plants. It's all done by cheap foreign labour. If that's too hard to automate, I'm not yet holding my breath for general purpose household robots either.

Admittedly the videos in this article do seem promising though, would love to see how this tech would perform in a greenhouse.

Oh yeah, the exact same way roller coasters got banned as soon as one failed and killed some people.

Laundromats are kind of a drag since you have to hang around nearby for almost 2 hours to do the 5 minutes of labor that the machines don't do for you. Dropping off a sack of clothes for wash & fold on the other hand is a pricey luxury. Perhaps if the laundromats could automate the whole process it would bring down the price of wash and fold and fewer people would be inclined to have washing machine at home (as you say, taking up space for a machine that is used a couple hours a week)

Alternatively you can simply attach the laundromat to a cafe or bar, turning a chore into an opportunity to relax and socialize but most of the world is not prepared for that degree of civilization.

If you scroll down, this is a general-purpose robot. It can drive around and bus the table or fold a cardboard box.

I'm not sure it can't handle the softener, and V2 will likely be able to set up the ironing board.

In terms of size, even now, it's smaller than a fridge, washing machine, dishwasher, or many other household time-saving appliances common in most houses (although not necessarily historic cities with multi-century homes not designed for them). No effort has gone into shrinking it either; perhaps with clever engineering, it can be made much smaller if it moves out of the research prototype phase.

Another question, to me, is cost. Many robots like this run around $100k, and with good reason. Can this be brought down? I hope so.

Why would you think a laundry robot is something people who can't afford comfortable apartment with their own washing machine might use?

I think this might be eventually integrated with washing machine so that you buy one device where you put your dirty clothes in and take out clean and folded. It's not like the traditional washing machine is the expensive part that needs to be kept separate and in current form.

> The cost of rent is 30€ per square meter per month in Paris.

Only for big enough apartments. 20m² appartements do not go for under 850€, ie around 42€/m². Granted these flats are too small to even have a laundry machine, let alone a laundry folding robot.

Agreed, my girlfriend has a biochemistry degree and works in such a lab. It involves some physical work like collecting, preparing and inputting samples, doing basic maintenance of the machines, some analysis and then some administrative/secretarial work in feeding the results into a system or by phonecall to the doctor.

All of which can indeed be automated in my view.

I would say it really depends on where you are though. In the US it probably makes sense quite quickly. But she lives in a small EU country where salaries aren't high, and this is very much a junior position with a lot of students looking for such a job. Her position costs about 25k a year.

The NPV of a $1m investment with a 25k cashflow is negative at normal discount rates. Once you get to replace a $120k salary with a $1m robot, it does make sense.

Further I do still expect there to be some jobs in overseeing the robot (e.g. your average factory manager). That makes sense for large centralised production locations, because you can have 1 human job overseeing many robots. It doesn't necessarily make sense with many small decentralised production locations. And that's the nature of most labs, I believe due to the time sensitivity of a lot of lab work they need to be everywhere and close to customers. But maybe that will change.

Don't expect much. Even screwing a nut on a bolt is a huge problem for generic robot. Which means you will need a 'robot friendly' lab. Were all things can be done by a primitive robotic hand. Other options are making more capable hands, and completely robotic specialized labs. The first is most interesting and the way to go. When it happens it will open a lot of possibilities. Like "self-repairing" vehicles and planet stations. Just with 'technician' robot onboard.

This lab automation product already exists and the couple of startups in this space face the same headwinds as everyone else trying to sell hardware when the stuff people are essentially really paying for is software.

Also nobody’s academic lab is buying $1m startup lab equipment. A whole core for 30 research groups is buying something, but it will be a piece of equipment that directly leads to publishing, ie, something with history. That is why you don’t have the exact thing you are talking about, which exists, in labs.

Also worth mentioning that a lot of stuff in chemistry labs is toxic and even potentially deadly to humans, it would be a big win if these could be handled by robots instead.

Certainly. I have a friend working in this space (AI lab automation). In a few years he expects competition to be fierce, though.

Laundry is kind of the perfect demo for advanced motion planning systems. Fabric is, for all intents and purposes, completely intractable in classic motion planning paradigms; it's wildly non-rigid, which means that predicting its behavior is the domain of highly specialized and expensive dynamics simulators, it's nearly impossible to invert the problem to ask what motions would be required to produce a given result, and it's highly continuous and resistant to discretization even if you can predict it. You can't make the "folds have zero width" assumption you always see when reasoning about origami, for example. Clothing is extreme even for fabric, given that it's not only highly non-uniform but also fragile; every shirt is a different hideous bit of floppy topology covered in strange textures with complex and unpredictable local properties and it'll start popping stitches if you look at it funny. Ruffles, zippers, pockets, drawstrings, the list goes on. On top of that, laundry is something that everyone does so it's relatable and easy to set up in a lab, and humans can intuitively evaluate performance with a glance. Despite all the attention, nobody's been able to demonstrate convincing performance on it in like seventy years of work, which makes it a more difficult task than backflips or shooting hoops or loading a truck. All of that together means that, when you have a fancy new algorithm that can handle more than some blocks on a tabletop, you pretty much always point it at the laundry.

> I don't get it why focusing on clothes folding though, does anyone do laundry more often than once a week?

Folding laundry is not the end goal here. They chose it because it’s a very challenging thing for a robot to do, requiring great manual dexterity, planning, reacting to sensory inputs etc. In other words: if your robot can do your laundry it can probably be taught to do pretty much anything else around the house.

> does anyone do laundry more often than once a week?

Once you have a nonzero amount of kids, laundry becomes a continuous process will no defined start and end.

> Much more useful if it can do cooking since that's something you do at least once every day.

Thing is, cooking is something a lot of people find deeply rewarding and humanizing (I don't, but I realize I'm an outlier here). Meanwhile, I challenge you to find anyone who thinks doing laundry is a worthwhile use of their limited time on Earth.

> I don't get it why focusing on clothes folding though, does anyone do laundry more often than once a week? Much more useful if it can do cooking since that's something you do at least once every day.

Lots of people do 3-4 loads per week. Some people don't cook at home at all. There are 8 billion people out there.

> I don't get it why focusing on clothes folding though, does anyone do laundry more often than once a week? Much more useful if it can do cooking since that's something you do at least once every day.

Because the people on the Axiom are going to need clean jumpsuits?

Seriously, laundry folding is an example of a complex task that requires considerable dexterity and sophisticated object manipulation on the robot's part, but unlike cooking it's relatively low stakes: little harm is likely to be caused to the robot, the environment, or bystanding humans if the robot screws it up. So it makes a pretty good research task that, when solved for, will make impressive demos at trade events.

> does anyone do laundry more often than once a week?

Yes, at least 3x per week. Young kids, toddlers, babies, workout clothes, sports clothes, work clothes, etc.

> does anyone do laundry more often than once a week

I can tell you do not have children. We frequently run 2 loads per day...

Also, I challenge anyone to look at that video from 2:00 to 2:45 (which is 90 seconds IRL) and tell me with a straight face this has anything to do with "intelligence". I have seen human babies at 5 months old severely outperform this thing.

Small businesses like Cafes? They will hopefully get this to laundromats once this bot can iron, too. Not even once did I iron or fold my laundry back in university ...

I got myself this perfect clothes folding robot about 8 years ago when I decided I would not fold or iron clothes ever again. Two useless activities forever gone from my life. There's more things like this around the house but folding and ironing are 100% useless. I pick the clothes from the rack, they come kinda folded in half but I don't even try, and all get dumped into drawers. I never lack space because I'll buy clothes only upon throwing out clothes, so the bigger volume of the clothes doesn't make a difference in storage space either.

Send your favorite recipe to any of the local robot chefs near you to get the exact dish you want, every time.

would restaurants with static menus even be relevant at that point?

von Neumann, Theory of self-reproducing automata (1966) p.72 "Role of High Complication":

> An automaton can not be separated from the milieu to which it responds.

> The characteristics of a human for survival are well defined on the surface of the earth in its present state, though for most types of humans you must actually specialize the situation a little further than this. But it is meaningless to argue how the human would survive on the bottom of the ocean or in a temperature of 1000 degrees centigrade. Similarly, in discussing a computing machine it is meaningless to ask how fast or how slow it is, unless you specify what type of problems will be given to it.

Fast fashion clothing increasingly cannot be ironed. But also it’s increasingly hard to buy good quality clothes because of the influence of fast fashion. Even very expensive designers will have shit like unlined coats or thin sweaters. Part of this is that the supply chain for good quality clothing has been greatly reduced. Even if you wanted as a designer to acquire good quality fabric, you have to pay more for it than you used to.

My guess is that speed is not their first concern.

They are using a vision language model to generate the robot motions token by token. They are being bottlenecked by the inference of the VLM.

LLMs literally barely improved since gpt3

Yes because clearly this is how automation works. Never mind that we've improved efficiencies many orders of magnitude over the last century and people still work the exact same amount of hours.

We could automate the world and then we'd all be stuck rolling a cube up and hill and back down. We will work, no matter what, even in the face of post-scarcity utopia.

People should be free to enjoy life their way. Including if they want to be grandpa Joe from willy wonka.

It's not really a step change.

Saycan demonstrated similar functionality with PaLM 2 years ago (2.5 years first demo of cleaning up coke cans, etc):

https://www.youtube.com/watch?v=Ru23eWAQ6_E

Caveat - I am way out of my depth here as I focus on Software AI, but: I can't see any major breaks from SayCan looking at this information. When studying robotics, Degrees of Freedom/basic Arms and mechanisms (tension, grip, extension, types of joint etc) are a typical case study. However, I do not think there has been any practical progress (with the lack of data) in about 20 years now, besides increased refinement from materials science (e.g. detection of soft materials) allowing some better grip but not particularly better dynamic movement(e.g. walking remains badly solved compared to the efficiency/mechanism of humans).

It's certainly an expansion of software capabilities with respect to what Boston Dynamics typically highlights in their videos.

With that said, Boston Dynamics is likely to have their own version of this in the works if not already live. They're videos just tend to highlight advancements in both hardware and software that are likely to go viral and connect with non-technical viewers.

You can already see it in some of BD's latest videos.

The trouble Atlas had taking the car parts off the slotted shelf makes it seem like they're quite a bit further away, given the hand dexterity it takes to fold clothes.

Mwell, the robot picking out the "laundry" from a basket next to the table is slightly more than I've seen before so at least they managed to combine two tasks that would previously be considered distinct (although note the very lengthy attempt to pick a corner of a green shirt in the "3x video" further down the article, which I think is the most honest of the bunch).

I agree this is still a tech demo showing a prototype robot solving a toy task. As usual we're supposed to think "if it can do that then it should be able to do this" but it never works that way in reality.

Sure, what you said is somewhat true, but you have seemingly misunderstood one key aspect here. They simply do not have enough training data to fold all possible laundry. If there was a way to fine-tune the robot on your own clothes, then it would be able to do your laundry just fine.

You're under the mistaken impression that this is just a scripted demo built around learning directly from teleoperation as was the case with mobile aloha. The VLA model here is capable of more generalization than you are willing to admit.

No intention to mislead, but to fit their gist into an informing title. I was originally linking to X, describing their tweet with laundry vid, then changed to point to their source blog post, where their title doesn't fully convey the content.

So this title is firm name and blog post title, adding the "laundry" from their hero video under the title and featured video on their x.com post that links to this blog post, as well as the first featured/labeled "Laundry" task discussed in body (not to mention, answering HN commenters such as this, since "doing the laundry" is by now a meme):

> I’m still waiting for a neural net that can do my laundry. -- sheeshkebab 11 hours ago - https://news.ycombinator.com/item?id=42098153

The response to this, portrayed in both blog post and x.com feature videos, is in the third sentence of the blog post:

To paraphrase Moravec’s paradox, winning a game of chess or discovering a new drug represent “easy” problems for AI to solve, but folding a shirt or cleaning up a table requires solving some of the most difficult engineering problems ever conceived. To build AI systems that have the kind of physically situated versatility that people possess, we need a new approach — we need to make AI systems embodied so that they can acquire physical intelligence.

The synthesis is, this is about “doing your laundry”.

Finally, to take it out of air quotes, the article's first explicit and bold labeled text task description is:

LAUNDRY. We fine-tuned π0 to fold laundry, using either a mobile robot or a fixed pair of arms. The goal is to get the clothing into a neat stack. This task is exceptionally difficult for robots (...and some humans): while a single t-shirt laid flat on the table can sometimes be folded just by repeating a pre-scripted set of motions, a pile of tangled laundry can be crumpled in many different ways, so it is not enough to simply move the arms through the same motion. To our knowledge, no prior robot system has been demonstrated to perform this task at this level of complexity.

They hold they have demonstrated this robot doing the laundry.

I hate doing laundry. Everyone gets one task they get to abhor. Mine is laundry. If I could buy this in a few generations for less than $20k, I would, pinches and all.

Perfect is perfect. Good enough is good enough.