Physical Intelligence's first generalist policy AI can finally do your laundry
physicalintelligence.company
physicalintelligence.company
Give me a moment
1. All economics / markers of value are about human happiness / satisfaction - we claim it’s money but it’s only given a value by humans who want it.
2. As soon as people get rich enough they outsource their domestic tasks - hire a maid or a cook, or buy ready meals.
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.
4. So most households have lost 35 hours a week, and also still have same amount of domestic duties to do
5. As we can’t give everyone a maid we might be able to give everyone a robot maid.
6. Most innovations / technologies find their way into homes - from bricks to heating to electricity we invent it and eventually find a way to make our lives more comfortable - see the point about economics is just humans liking stuff
7. I assumed that real robo maids would be a social shift - ie a different design of washing machine, people eating at other peoples houses every day, anti-dust surfaces. But this one looks … interesting
8. I know this is incredibly western middle class centric - but exactly what else are 6 billon people aiming for?
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.
Is the solution really to replace even more workers by capital, or do we have an issue with how we measure value that we should fix first?
And yet countries that value more than efficiency and productivity are repeatedly ranked as the happiest in the world [0].
"Just push for progress in natural way, and for politicians to adjust to an already happend changes"
How much damage has unfettered tech done to society at large? Maybe we'll never know the full answer, but we can definitely agree it hasn't all been positive (advertising, social media etc..). We're now retrospectively regulating, but a lot of it feels too little, too late, for some elements of society (privacy being a notable example).
0: https://worldhappiness.report/ed/2020/the-nordic-exceptional...
I have more faith in our ability to solve world peace and AGI than I do in us getting to a more objective way to measure value that everyone can agree and adhere to.
In other words: we're screwed.
The problem with the wealth-weighted-value that gets optimized by capitalism is that the gini coefficient tips the optimization process from being about doing what other people want to being about doing what rich people want. Rich people mostly want to get paid for being rich, of course, so they pump assets to increase their wealth. Their weight goes up, the objective function pumps assets harder, their weight goes up, the objective function pump assets harder... and gini heads to 1 and you return to a palace economy.
A note of optimism: we've been here before, shortly after the industrial revolution. We've fixed this before, even though Marx predicted that we couldn't. We should all be trying to figure out how to make sure that next time the USA gets neo-Roosevelt, not neo-Hitler or neo-Lenin.
Is it fixed if it falls apart in a couple generations? Maybe flex tape over the problem isn't going to cut it.
On the wealth weighting, I’m a fan of a 100% inheritance tax tbh. I’m in the “top 1%” and I don’t plan on giving my kids more than a great upbringing and education. I’m not giving them any additional cash injection. I think that solves a lot of the wealth weighting problem by shortening the lifecycle of capital holds, but I don’t know if I’ll ever live to see it (or some form of it like inheritance caps).
We do?
But if you also buy clothes only made from hand-spun yarn, I mean, it was a problem for the yarn-spinsters back then, and by all means help them. But paying someone to do something that no longer needs doing is how you create bullshit jobs. No problem with spinning yarn for recreation, but if they have to do some complete nonsense to earn a roof and a meal... there are kinder ways to help.
Alternative take wrt. self-service checkouts: the store is making you do unpaid labor for it. And with how it's implemented in practice, it's not only offloading work of a specialist to you, it's also costing you time and frustration, since a cashier doing checkout 8h/day on a checkout desk optimized for throughput is doing a much faster job at it than you doing it ad-hoc, in constrained space, on a machine that throws an error if you look at it funny.
I'm very much for automating bullshit jobs away, but store checkout ain't it yet, and the results are universally worse.
More importantly, it's a common pattern that's been inflicted on people (and ourselves) by our industry particularly often: "automating away" specialist jobs with software. The workload doesn't actually go away - it gets redistributed to everyone. It only looks like a win, because specialist salaries are legible to the business, while generalized productivity loss isn't.
Think of it next time when a two day coding problem takes you a week, because you also need to attend several useless meetings, make some powerpoint slides, update your calendar, timesheet, and fill in expense reports from a business trip.
From my perspective self checkout is a latency optimization. The person in front of you in line may have 3 items and still take an unreasonably long time finishing the transaction, but replace one line with 4 machines and the line keeps moving.
Sure I’ll take a full cart to an actual cashier, but by removing people with only a few items from those lines they become more predictable.
It doesn't look this way from customer perspective, so I propose alternative interpretation: it contributes to hiding inflation, because instead of seeing the prices rapidly rise, we see them rise slower (further slowed down by "shrinkflation" and related shenanigans), but the quality of service goes down too. So what we get is ever worse shopping experience and lower inflation rates reported by economists.
(Not to mention, shopping taking more time because of this "optimization" is effectively even more hidden inflation.)
> From my perspective self checkout is a latency optimization. The person in front of you in line may have 3 items and still take an unreasonably long time finishing the transaction, but replace one line with 4 machines and the line keeps moving.
It's a good reason to have both. Because in practice, 2 out of 4 machines are, at any given moment, locked and waiting for a supervisor - and many stores "optimize" further, by eschewing a dedicated role and just tacking responsibilities on to existing work roster. Meaning, the supervisor you're waiting for is likely staffing a checkout point or unloading boxes in the back of the store at the same time, so you better get ready to wait.
Ideally, the store would have both automated and staffed checkout, and direct those with few discrete items to the former, where a dedicated supervisor would ensure problems get resolved quickly. That would indeed optimize latency/maximize throughput.
I’ve never been to a store that only had self checkout, I think the equilibrium is one bank of self checkout + a bunch of normal registers.
> 2 out of 4 machines are, at any given moment, locked and waiting for a supervisor > where a dedicated supervisor would ensure problems get resolved quickly
Sounds like an implementation issue.
The grocery store I typically use has one bank of 6 self checkout machines with a dedicated attendant who seems to be able to resolve any issues. It’s common for 1 or 2 of the machines be down though. They also have 4 to 14 normal checkout lines open depending on how busy the store is.
I rarely use self check but it seems like a good tradeoff vs a 5th line being open when things aren’t that busy.
Yes! I have been thinking this. So as an economic consequence of both parents being able to work, now the choice to stay home and care for the kids is an even harder one. The birth rate falls. Kids are raised by childcare workers, not their parents.
We need to stop asking 10+ workdays a week from each couple. Once we only demand 7 workdays a week, people are allowed to raise their own kids again!
Turns out, for the market, having a choice is a form of wealth. It will make you pay to acquire it, it will treat you as that much more wealthy if you have it, and it will encourage you to sell it.
Ultimately capitalism is built so that improved productivity, more working hours, higher salaries never benefit the worker long-term. If people start earning more, the owning class will simply raise prices to match this new level of income. The US is the wealthiest and strongest country in the world, and yet young people there struggle to buy houses just as much (or even more) than their peers in much poorer countries.
All sexually reproducing organisms on the Earth have families without any social pressure and we evolved from primates who work like this.
By far the stronger pressure is biological. I doubt social pressure even comes close to moving the needle.
The peak reproductive years are the late teens to late twenties (give or take). However and in modern society, having children in the first two-thirds of the peak years is a recipe for failure or lifelong hardships.
Is this bad?
The thing you're missing is that, you go to a starbucks, which promises to sell you posh coffee, but it may be good, or it may be crap, and it's all down to luck whether the barista gives a crap about your "posh" coffee. Whereas if you go to a small business owner whose business is posh coffee, you can probably expect a great coffee (or if they don't, then you expect the place to soon go out of business)
hardly requires a bustling industry to solve this 'problem' or provide this 'value'.
starbucks is a monument to the lazyness and exploitation of this lazyness in homo sapiens.
The only 'skills' you need is the ability to transfer powders and liquids from one container to another.
I have no idea why people idealize coffee making so much. I have been doing pour over coffee my whole life and it feels as basic a skill as washing a dish.
I have a "hobbyist conjecture" that I'd love feedback on from people familiar with economics. Simply put, the idea is this: if you were to give every human around the world a million dollars (imagine this is easy and feasible), prices would inevitably rise. The poorest might experience some improvement in their standard of living, but not as much as one might expect. This effect seems to go beyond inflation alone. My intuition is that prices ultimately adjust based on social structures and expectations.
Similarly, we see a unique economic advantage in many dual-income gay couples who, statistically, achieve higher standards of living. Often, this is partly because fewer have children, which shifts financial priorities and spending capacities.
Other currencies 1M base notes is not a lot (e.g. 1M dinar). You can just add/remove zeros but prices have adjusted, those people can't live like "millionaires" with 1M dinar.
There was a time goods like meat cost pennies, now it's $10 per pound. In those times $10,000 would be a life-altering amount of money, today most people have $10,000 in assets. The price of goods is related to money supply, they get more expensive if people have more money.
Money has no intrinsic value, it is balanced by whatever goods and services can be bought by it. If you add money but no goods and services, money is worth less (see COVID policy, increase money and decrease goods and services).
More generally, gifting every person worldwide the same amount of money seems roughly equivalent to taxing every above-average wealth person a fixed percentage of their surplus and giving every below-average-wealth person a fixed percentage of their deficit...
On a more realistic note, gifting every human 1M would probably completely break the financial system and cause a global recession...
So, if you give everyone in a given population (I'd go for the country instead of the world to see the effect much faster) a million dollars a month, the average discretionary income would, obviously, rise by one million dollars. That would cause prices to raise, but also all kinds of new products and services to become available. Some of those would evolve from luxuries to necessities over time - like e.g. having a car, or a mobile phone. Housing would likely eat a chunk of the surplus.
In the end, it would all stabilize, at exactly the "life is more expensive by a million dollars, on average" point. We'd still have socioeconomic classes - poor with < 0 surplus, rich with > 0 surplus, and the middle with ~0 surplus. Some people would've moved to different places on the ladder, but the ladder would still be there, simply because of variability in individual incomes.
I guess put another way, you could say, the market is a DC filter - as wealth distribution in a population changes, the market always adjusts to cancel out the constant component ("DC offset").
Changing policies frequently increase the risk for future arbitrary decisions.
Doesn’t matter whether a person chooses self check out or cashier check out.
What matters is the training.
When self checkout becomes nearly as accurate as cashiers the cashiers lose their jobs.
Which lane we choose has little to do with the replacement. Thousands of holdouts won’t affect the millions of shoppers training the replacements.
Note: The European guest worker model of the 1950s and 1960s doesn’t work for the United States, due to the birthright citizenship status.
I’m not sure this is Owning the Bots the way you think it is.
This is just not true. If people feel worse it's because they want more things from life, not because they are able to afford less things.
(I'm talking about the last ~50 years. The last 20 years may be a different point, but it doesn't have to do with female participation in the marketplace)
Not sure how that shows up in the stats
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)
https://www.cbc.ca/news/politics/canada-gdp-per-capita-rich-...
https://nationalpost.com/news/canada/canada-immigration-per-...
The first thing that robot will do is start a dryer that a toddler climbed into because it isn't that aware of the world around it.
And that will be the end of general purpose domestic robots.
That or knocking over candles or fucking up something else simultaneously trivial and terribly dangerous in context.
I dream the same dream of a general purpose machine, but I think it may never be possible, and if it is we're a long way out.
Anyway, if we're going to fake smart robots this way, why not just honestly call it what it is? Remote household staff. Might as well give them better sensors and actuators, but I guess this is giving humans too much agency and risk; what if the operator decides to hurt the "robot" owners or something? The vendor would not have that. Cannot have that. Humans are too messy to deal with.
Or is this the short-term future of all automation? I look at my robotic vacuum cleaner now, occasionally pausing for a second or two while it figures out where to go, and I wonder - maybe in those moments, it's using some Protein Intelligence Chip to query a bunch of random humans somewhere?
They are actually controlled by "robot operators" in Estonia, though officially it's "AI".
That being said, every day it's getting a little bit better.
Estonians are fast learners!
I think that will weigh on people's opinions of domestic general purpose robots when compared to robotic cars.
On the other hand, virtually no one dies from laundry, that I’m aware of. So the reaction to a single accident might be quite different.
Like that old quote: "I love work, I can watch it all day."
Admittedly the videos in this article do seem promising though, would love to see how this tech would perform in a greenhouse.
While they aren't widespread, there are production robots being used in many types of agriculture. In fact, it was even trivial to find one working in tomato greenhouses: https://www.arugga.com/technology (this one is used to pollinate tomatoes instead of doing it by hand or using bees).
Living in European city, space is a hard constraint. The cost of rent is 30€ per square meter per month in Paris.
Laundry rooms are small. This robot is too wide and won't be able to go through the door of my laundry room. Ironing boards are foldable for a reason : they need to be setup every time. This robot can't do it, and also can't handle the softener bottle for the washing machine.
Having 1 square meter empty table (0.5 for the table and 0.5 accessible space for the robot doing the folding) dedicated to folding is a pipe dream for most. Laundromats are there because some don't have enough space to even have a washing machine.
Laundry room are a dedicated space for humidity and ventilation reason, so they have been designed on specific location on house plans probably more than 30 years ago on average, not having in mind robot accessibility, but rather be as small as functionally possible.
Quite often for people not living in flats, but in houses, the laundry room is located in the basement with only stair or single step access.
I don't think architect and construction accessibility norm will change fast enough, specially with bipedal robots right around the corner.
The slack necessary for home robotics emergence has already been eaten multiple time due to the high cost of space.
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.
Ironing board is bigger than the robot, and quite heavy, the torque in setting it up is a lot and probably greater than the torque necessary to break the expensive actuator when things go wrong.
Keeping a shared table clean so you can fold clothes on it is a not trivial but more manageable problem.
The home market is highly predatory. If people have enough to invest in a home robot, it means that they have some available money that other actors didn't grab. In France last time they remove 5€ per month of renting help, it was a struggle for lot of people.
You have lots of actors that call dibs on your home improvement money before you can improve your home. There is usually the landlord, then the building manager (or Home Owner Association) (which is usually already working in the kleptocracy domain), then there are local tax, and mandatory ecological improvements (thermal isolation, and windows), home improvement market (paints and kitchen), and of course everything is indexed on the current cost of energy, and uncontrolled inflation.
It's not a question of costs, otherwise all smart investment improvements that bring more than they cost on the relevant time-scales like thermal isolation and solar power infrastructure, collective infrastructure, not having to do the laundry every week..., would have been done already. It's more about determining who deserve the money.
Actuators, competently designed, won't break when torque gets too high:
1) The control system can handle this, trivially.
2) The mechanical linkage can be designed to limit torque. At some point, it should slip.
3) Forces out-of-axis should not be transferred to the actuator
The kleptocracy is more a function of rent-seeking city living. Even a little ways out, into the suburbs, things are very different. Rural, and space is virtually free.
The mechanical part of robotics is quite challenging. Because of the overhangs, 6DOF arms usually are bulky to be able provide useful force. Alternatively too strong industrial robots are unsafe to operate near humans even if compliant.
But most probably even robot building startup can't even attempt it. (They are not looking for mechanical engineer in their job listing, so probably they don't design the robot arm themselves).
If you don't design the robot arm yourself, then you only have the option of which off the shelf actuators to pick. Usually they are weak and expensive. And how to integrate them while staying in specs.
For manipulation sensory input is very useful, but if the API of your robot arm doesn't provide the feedback you do with what you have. When forces are weak closed-controlled loop system don't flex and can know their position, when things flex and deform, it allows to have lighter arms, but the robot can't know it's internal state. Here they close the loop via visual feedback, so knowing position exactly is not so important : humans have shitty repeatability but handle folding the laundry quite well.
But shitty robot arms are not very useful and also need the same motors and controls that good robot arms, so you might as well make your task easier and develop your robot with expensive arms first so prices converge towards expensive robot arms, and the economics equations of this market (high R&D cost, low sell volume) are geared not for home applications but industrial control which usually necessitate maintenance operations on the robots themselves.
What's even worse for the business model, is that once you have spend money to gather data and build a dataset to teach your neural networks, then your competitors only have to copy, vertically integrate, improve, scale and grab the money pot.
One of home robotics main obstacle is a https://en.wikipedia.org/wiki/Collective_action_problem where no one wants to front the costs of building the datasets, front the costs of cheap actuators, front the costs of cheap prototypes so everyone reinvents their own solution which can't solve the economic equation.
When I was younger, I'd have argued the collective action problem is just a matter of time and development, but I've now seen dozens of technologies stalled out.
I firmly believe it's the reason we don't have flying cars. I can sketch out all the pieces needed to get there, which in the simplest-to-explain version consist of a few regulatory changes, an autogyro, a control system, and some infrastructure, but those few changes easily add up to an obscene amount of money. Once invested, the next competitor can do almost the same thing almost out of a garage.
Lots of medical technologies too. And medical systemic changes. Education too.
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.
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.
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.
An adjacent use case is apartment buildings that currently have shared washer and dryer rooms that tenants book time slots for. Upgrade: drop off a tagged laundry bag anytime, pick up when convenient from locker. Only service technicians need room access so layout and washers and dryers can be robot optimized.
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.
Breakfast and lunch are not paid breaks in any EU country as far as I know.
One 30 minute break (lunch break) is paid in Croatia and Slovenia
Most office workers spend an exorbitant amount of time twiddling their thumbs, reading emails, going to the restroom. On average, I'd say knowledge workers are typically productive for about 2 to 3 hours per day. I'd estimate physical laborers range up to 4 or 5 hours.
Assuming 8 hours of full productivity is a strange number to focus on
Again for large scale centralised labs that get sent samples throughout the 24 hour period from all across the world, where timing isn't an big issue, this could work. But most labs are small, close to the customer, time-sensitive, and work with couriers to bring samples because samples need to be transported and stored in specific ways not to go bad.
She works quite long shifts btw (10-12h), but that's more a function of her country/company culture rather than the norm. Probably 6-8 is more common indeed.
I could see demand spreading out more - though I agree it might be closer to 2 than 4 - if the availability of human labor wasn't an issue.
Just a bit lower than what a human would cost.
[1]https://docs.google.com/document/d/1_3yhWjodSNNYlpxkRCPIlvIA...
[2]https://mobile-aloha.github.io/
$7000 for the wheels/navigation module $3500 x2 for the control arms $6000 x2 for the controlled arms
I don't think you need a $7000 navigation module and the arms are also probably overkill. To which extent do we need closed-loop Dynamixel servos when the policy diffusion algo already relies on computer vision driven closed-loop control ?
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.
Instead I would like to see some new innovation in laundry machines. Current technology is very basic, you heat water and keep spinning. Why not some something more similar to bigger paper printer that at the end of the day you feed your 1 shirt, roll it into a roller, sprinkle with some high pressure minimal amount of water, heating and return 30min later already perfectly ironed.
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.
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.
Yes, at least 3x per week. Young kids, toddlers, babies, workout clothes, sports clothes, work clothes, etc.
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 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.
The big advance here seems to be that the robot can pick the clothes out of a basket on its own rather than having someone set it all up neatly for it. I mean they sort of imply it here but you have to read carefully to understand what they refer to (folding a t-shirt that hasn't been laid flat on a table first):
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.
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.
Pretty much this. You’ll be folding clothes while washing more at the same time.
And it isn’t just the folding that is a challenge. It’s putting it away. I’d gladly take a “put the clothes away” robot before I got a “fold the clothes” robot.
This also means I use less floor space as I don't need cabinets/dressers to hold them.
And yes, I have many racks on my walls. The rack life is good.
would restaurants with static menus even be relevant at that point?
I never thought about what we could do if instead of trying to do work in real time we slowed robots down for non time critical tasks to latencies which current transformers can deal with on real hardware.
> 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.
I think "doing the laundry" is getting obsolete faster than doing it yourself is. Kind of like people didn't give up on sewing their damages clothes by hiring a robot to do it... they gave up by throwing old clothes away and switching to fast fashion.
If this product addresses high-income people (which would probably be the case even if the price comes down from $100k to even $1k), then it still seems a lost cause for this reason.
Or maybe we will have private army robots under the order of some random dictator disconnected from the reality of people.
Or some mix of these. Or finally no longer any war, in a world without any human though.
What a time to be alive.
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.
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).
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.
- changing a bike chain
- closing a jacket's zipper
They have an extremely specific definition of the word "laundry". Three specific shirts of size M, L, and XL, and two pairs of shorts of size 28 and 36. Then they do a mobile task on three specific shirts of size M, M, and XL, and two shorts of size 32 and 31W.
That's not "laundry" the way any lay person would understand.
It's sad to see yet another group hype their work for money. What they did do, is invest a tremendous amount of engineering effort to scale things up, to collect data, and to see what changes models need to scale well in this space. That's nice and useful. It's something that as academics we generally cannot afford to do. And won't be able to afford any time soon because they don't share their data and none of their results are reproducible.
PS: Where is this title from? "Physical Intelligence's first generalist policy AI can finally do your laundry " It's not in the article. It's very misleading.
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.
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 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.
Perfect is perfect. Good enough is good enough.
In my neck of the woods it costs a couple years of the cleaner dropping in every week or two. In New York, maybe a couple more years, albeit with the added burden of having to pick it up and drop it off myself.
We're excited to show the first step towards this mission - our first generalist model π₀
Paper, blog, uncut videos: http://physicalintelligence.company/blog/pi0