Obviously, this type of “robot math” depends entirely on the cost values you assume. If a humanoid robot only costs 5x the cost of a specialized robot, then of course you’ll get the outcome that the humanoid robot is better. But that’s totally inconsistent with what we’re likely to see.
Humanoid robots capable of doing multiple versatile household tasks aren’t going to cost $500. To use Optimus as an example, Musk is aiming at a $20K production cost, so assume somewhere between $25-$30K retail cost at scale (and Tesla is not known for hitting their price figures). Meanwhile, the most ubiquitous specialized household robot - the lowly robot vacuum cleaner - is around $250-300 for a decent (but not luxe) model. That’s a 100x price difference, not a 5x difference.
Hey, I agree with your analysis - but, we have no examples of that scenario actually being factual when it comes to human form functions - specifically our method of movement.
Let’s use Tesla as an example. They have a larger number of very expensive robots in their factories building cars:
Many of these robots can switch a tool on the end of their arm so they can perform multiple functions - but they are stationary. They don’t have legs moving them around the vehicle. Can this process be improved? Very likely. Will adding legs and feet be the improvement? Not likely.
Does having a phone that can double as a music player and a home computer increase function? Absolutely. Would adding legs and feet to it help? Doubtful.
Not suggesting it can’t be done, just that full mobility is a challenge, especially since it has to be achieved in a package suitable for what else the device is supposed to do.
There are certainly contexts in which a single purpose robot is the right idea, but I think the focus in this conversation is or should be multi-functional, highly multi-functional. A robot whose only skill is making the bed is not going to be a big seller.
For sure. Given the state of the art (videos I’ve seen, and projecting improvement into the future) I’m just sort of wondering what kind of thing they will be able to do, at least short term.
Fold the laundry.
Make the bed.
Vacuum the floor.
Clean the windows?
Fetch the mail?
Walk the dog.
Sew a patch? I don’t think so. Those needles are tiny and the dexterity required pretty amazing.
I must say I’m struggling to think of many more. It’s not going to climb a ladder and clean the gutters, or use a power washer to soap up the car, is it?
We won’t need it to drive the car, as that will drive itself, presumably.
Could some of the advocates add to the list, and perhaps that would give me an idea of how important “legs” are.
Cook a meal? I would think it would be a while before it could get into my closet/pantry, pull out a cereal box, dispense the right amount of flakes, add sugar, and have the milk ready for me at the table, but perhaps I’m too pessimistic - or more to the point, too cheap to spend $30k for the privilege.
Consider the robot vacuum cleaner, though, which only does one thing. They’ve sold a lot of those. I’ve seen a few robot lawnmowers, which also do only the one thing. Google tells me that about a billion dollars of those are sold every year.
I don’t think anyone’s buying a robot to just make the bed. But the other big household chores are: 1) laundry; 2) dishes; and 3) cooking. I don’t think #3 is on the cards for either a single-purpose or a multi-purpose domestic robot. But I can see people being willing to buy a single-purpose robot that could do laundry and nothing else (from hamper to fold/hang, if not put away) for the right price. Same for completely doing the dishes.
I’m not an advocate, but I think one idea is that the robot would go beyond the “housework” box and more into the personal valet/butler arena. So you’d order your groceries from Instacart (or what have you), and they could be delivered when you weren’t home - and your robot would bring them in from the street/stoop, and you’d come from work and your fridge and pantry would be fully stocked. Or you’d send your RoboJeeves to the store, or to pick up your dry cleaning, or to get your take-out order in your self-driving car. Or the store or drycleaner or UberEats would have a bunch of these.
They’d be the “last mile” type service to cross the gap from the robot car parked downstairs, that gets things into your apartment or home - across the two steps up your stoop, through your door, and opens the fridge/cabinet/closet to put things away. I also wonder how big the market is for that: $30K financed over five years at 6% is close to $600 a month for that kind of robo-service. But there might be some people who would be willing to pay for that.
Yes, but if someone was significantly incapacitated, then $30K might be cheap.
Right, but it doesn’t cost $3000 either. Plus it does a task in the background which many hate taking time for. There is a whole spectrum here of cost and perceived value.
People that are significantly handicapped often cant afford a wheelchair, must less a $30k robot. But let’s assume every single person that uses a wheelchair will buy a $30k robot. That number is estimated to be about 2% of all Americans. Not a robust market for long term care robots.
I wholly agree - but cheap is kind of the point. A lot of the time, simpler things are cheaper. Amazon uses the types of robots they use because they are simple and cheap, with the number of parts slashed with each generation/iteration to get initial costs and maintenance costs down.
A single humanoid robot that can sweep a floor and mow your lawn and do the laundry would be a technological marvel….and perhaps far too expensive a way to do all those things compared to alternatives. Every time I see the (annoyingly) large numbers of people riding scooters around downtown, it reminds me of Dean Kamen and the Segway - an engineering marvel of a solution that didn’t make it, compared to the much less elegant/functional but much cheaper substitute that did.
Ironically, it’s the Innovator’s Dilemma scenario for a product that isn’t even out yet (the humanoid robot). If most people will never need 95% of the things a humanoid robot is capable of doing, then it’s really vulnerable to having its market niche occupied by cheaper “less good” robots that aren’t nearly as wonderful, but have the advantage of being right-priced and right-functioned for what the consumer can afford and needs.
Not a good example, obviously these giant robots are in a different league. Humanoid robots are by definition for jobs that a human would be doing because there is no other better alternatives. Also one must consider the pace of adoption. In time these robots will interact with humans but initially they will be used in factories and businesses. The test case is in Tesla factories.
You might have missed it but personal computers happened several decades after office computers happened. Their price dropped from millions to thousands to hundreds. Robot prices won’t drop quite as radically but drop they will.
I figured a cost of $10K so let’s use your $30K retail cost.
How much does a low level Tesla worker make? At $25 an hour at 40 hour work week adds up to $52,000 a year. Add 30% benefits, $52,000 + $15,600 = $67,600 a year.
A 40 hour work week adds up to 2,080 hours per year. A robot could work two shifts six days a week which adds up to 4,992 hours, 2.4 times a human worker [no stock options!]
2.4 workers cost $67,600 * 2.4 = $162,240 per year. The $30K robot pays for itself in 67 days. Add robot expenses that might make it in 90 days. Useful lifetime two or three years…
And completely unnecessary. It’s very difficult to balance a top-heavy humanoid. Move to a quadruped, and it can do everything you just mentioned. Or even a wheeled robot. Making it humanoid doesn’t really add anything to its usefulness.
Perhaps. It’s not like robots are themselves new things. Like your “office robots,” “factory robots” happened many decades ago - and robot prices haven’t really dropped the way that computer prices have dropped. Office computer prices went from millions to thousands to hundreds as well, but we haven’t seen the same thing happen with industrial robots. So perhaps robots won’t exhibit the same price curve as computers?
If an Optimus could actually replace a Tesla factory worked, then sure - that would be a very valuable product. But we’re a very, very long way from Optimus being able to replace an auto factory line worker whose job has not been (and therefore cannot be) already replaced by a robot.
The problem is the brain, not the body. And the problem is data. Unlike LLM’s, which have virtually the entirety of everything ever written and published in English (mostly) already digitized and published and scraped in the Common Crawl, there’s no vast repository of data on humans manipulating objects in the world. There’s third-person video, of course - but that’s not going to enable a robot to get data on force and torque and pressure and whatnot for doing things like opening a door or picking up a glass with liquid. Much less how to perform an auto factory job that’s managed to elude mechanization for five or six decades.
If you look at the below video, one company is trying to solve that by having early adopters of their home domestic robot allow teleoperation. Basically, the company acquires data by recording what the robot does during teleoperation (which it can then use to train its model) in people’s homes. So if you do this for many thousands and thousands of hours, you might build up a training data set that could let robots start to learn how to perform household tasks.
But they’re going to be slow and make a ton of mistakes in the teleoperating phase, and in the training phase - and they’re not likely to be as good or sharp as humans for many, many years. And that works for home robots. If it takes my home robot 67% more time to do the dishes than I would, I don’t care one whit. If it makes more mistakes than I would, that’s mildly annoying, but tolerable. But that can’t work for an automotive line. If you replace a worker on an auto line in one of Tesla’s plants, and they’re 67% slower or making a materially higher rate of mistakes, it will cost you many many more dollars than you’re saving in salary and bennies.
Neither are lawyers - I never claimed to be an innovator.
Many, many ostensible innovations don’t succeed. Many others don’t succeed until a very long time after they start being developed. I’m not in any way arguing that we shouldn’t try to come up with as many innovations as we can. I’m just skeptical that the technology has advanced sufficiently such that we’re only a few years away from having the type of humanoid robot that would be able to replace workers in Tesla’s factories.
The technique we’ve used to train LLM AI and self-driving cars involves massive amounts of data, which data was relatively easy to collect. We put a gazillion words and images on the internet, available to the Common Crawl. And it’s relatively straightforward to put cameras (and LIDAR and radar for those that use them) on cars and collect a ton of data by driving around.
Data collection for robots is much, much harder - because it probably can’t be done passively like that. Instead, you’ll have to create it by hand, with haptic gloves and teleoperators just doing routine stuff over and over again:
That’s 1X’s approach (in the WSJ article). Tesla apparently is trying to move away from this and do video-only - rather than collect training data from the robot’s perspective, they’re going to do just tons of cameras from all angles and try to get the robot to learn from just video. That can’t possibly teach the robot enough about things like force and friction and torque and weight and how to hold imbalanced objects and the like for it to be able to manipulate things in its environment - but it might be enough to get it to learn how to walk around an investors’ office or do kung foo moves. But that’s not going to get it anywhere close to being able to take a job on an assembly line.
That’s exactly my point. No, I would not. So if I had a system that was proven to work for training poets, I would not assume that the system could effectively train a blue collar worker. Having a method for “building” an AI poet does not mean we have a method for “building” a blue collar worker.
But not for factory workers. That’s the point. You can use robots being teleoperated at home to learn things about the objects they encounter in the home environment - how much a 1/3 full half-gallon jug of milk weighs, how smooth (friction) a dinner plate is, how much force it takes to open a kitchen cabinet. But that won’t ‘translate’ into that robot knowing how to properly manipulate a wire harness in a car chassis.
Some bots are just audio, not physical, much as the supposed Xfinity ‘Agent’ I talked to/with, this morning… Almost lifelike, but overly courteous, and voice patterns that were built in… In the end, it accomplished the task, explains away the Paramount+/Showtime on my bill is not really the Paramount+ App that I’d been trying to log into, no, it’s a separate subscription, so I gave up… No proof it was a bit, but… Customer servie just isn’t what it used to be…
