Why humanoid robots won't catch up to human workers any time soon(understandingai.org) |
Why humanoid robots won't catch up to human workers any time soon(understandingai.org) |
These humanoid robots (or generally limbed robots) could be useful for the two sectors that are craving for cheap workers, worldwide: agriculture and construction. Yet all the demos i see are fragile lightweight robots not made for rugged work. That makes me conclude that the whole sector is currently like the Metaverse.
I think the same will happen with LLMs. They're going to create a vacuum in the economy, on a global level. And I think what will come next is something in the real world (as opposed to the digital) filling it. Humanoid robots might be myopic, but I think it's, at worst, a pretty decent guess. One basic goal would be able to automate typical service jobs, both in terms of normal employment - but also even within households. A low cost highly capable robotic maid would have a massive market.
I'm hand-waving away the dystopia of this all, because I ultimately think humans will still play a key role - but trying to predict what that role may be is impossible from where we're standing. Imagine trying to predict the consequences of mass industrialization when you've lived you're entire life in an agrarian society. There's just no real basis for an intelligent prediction, besides 'way different.'
There are plenty of YouTube videos available were you can see how many hands a rice cooker is going through.
And we talk about a real robot costing 5 figures being able to shown what it should do and then it just does it 24/7.
It will fill the gap of no automatisation to full automatisation.
Besides that, in theory it would also allow the remote-gig economy: A robot at home, a helper sitting somewere else.
Take some random problem: You want some automatistaion for painting a wall. What do you do? You buy the full automat from a special company which comes with its own support contract, and fllow up costs.
Now you buy one 5 figure robot, explain to them that they should paint the room and they will do it. Not today, but do you think this will take 50 years and is a non solvable problem? I don't.
There is , however, real shortage of humans for agri and construction , in the rich world, right now.
> You want some automatistaion for painting a wall
You call over someone with a spraying machine that paints the house in 30 minutes.
I feel so many people bullish on these robots has never worked on physical things so they have this rosy picture of the smooth sailing once they get their humanoid robots delivered that it's going to just work forever without effort or input.
Our entire world is heavily adjusted towards bipedal humanoid creatures. Having a moving humanoid tool who could replace some of our physical tasks without the need to build additional infrastructure in proximity around it is BIG. Noting your reservation about fragility, even the possibility of cheap construction and agriculture work is VERY BIG.
Bipedal is the easiest one: very few environments require 2 feet in particular. Sure, ability to cross uneven terrain is needed, but this can be done with 3, 4 or 8 feet. But in a lot of places, including most of the retail and a lot of factories, all one needs is an ability to move on flat surface and cross door thresholds (thanks to ADA, no stair climbing required). And than can work just fine with large wheels, or tracks. On the other hand with something like construction, you probably want many appendages with some sort of climbing attachments?
The "top" part of humanoid is more integrated into the world: our human infrastructure is really built for beings just under 2m high, and a many of actions need at least 2 human-like hands, with at least the same reach as humans (~3/8 body height long, attached at ~7/8 body height). But this is just a minimum - a lot of things would work better with 3 hands, or maybe with dissimilar hands - one "strong", one "fine manipulation" and one "misc"? I certainly wished to have a 3rd hand when building something...
could any of this be reversed?
Unless Americans are willing to work for the same wages, the same hours and under the same conditions, no. Most of those jobs have already been replaced with automation anyway.
The only sector I see any of these selling in any meaningful number is military and police.
I think it’s probably also easier to raise money on a humanoid robot because everyone has some intuitive understanding of the form.
I think for safety reasons, we won't want any home robots to be very strong. Just precise.
A better obvious use case for fully humanoid is vehicles. Jump into a forklift, combine, or golf cart and be immediately productive. Then jump out and a human can use the vehicle still.
The physical world is currently designed for human interaction, and people who want to make money see the potential to replace humans with property. They'll get there.
"Your scientists were thinking so much about whether they could, they didn't stop to ask if they should".
edit: realised there's a part 2: https://generalrobots.substack.com/p/benjies-humanoid-olympi...
The real reason it’s hard to assess where we’re at with industrial robots is that some of the people who’ve gone hardest at it are secretive, and also believe in power law scaling. So, the first 1,000 robots will only be used to build the next robot factory, which also will be used to build the next really big robot factory. When we see humanoid robots actually out in the wild for sale at retail, they are going to come very very quickly, at scale.
Figure and Boston Dynamics sold out next year’s production runs almost instantly sometime last year; I think it’s just incredibly hard to overestimate the impact humanoid robots are going to have on the global economy.
All the dancing, jumping, running, etc. is nice for marketing but not actually that urgent for such tasks. Dexterity is nice. There are plenty of examples of humanoid robots doing pretty fiddly stuff like folding t-shirts and other things.
The rest is just optimization where these things gradually get faster, efficient, do more complex things, etc.
Depends on the context. If they are to work around human beings, particularly in the home, then they must be safe... to work around humans.
If a humanoid robot hits a human being, knocks one over, or even falls down on top of one -- especially a child, who is more fragile and whose movements are unpredictable -- then the industry as a whole faces a huge setback. Just look at what happened when a "self-driving" Uber ran over and killed a pedestrian in Arizona.
The different challenges presented in the article are almost all orthogonal to each other. Some of them are hardware engineering challenges, some of them are software / AI challenges, and most of them ultimately economic challenges. They can all be solved in a mostly independent manner, meaning that they won't be decades away to be resolved, but more likely years.
I think robotics is in a space that very much falls into the "people overestimate the progress in one year but underestimate the progress over 5 years" category.
A lot of the problems to solve regarding effectiveness (not efficiency) of humanoid robots even as outlined in the article (manipulation = short term spatial planning; long term task planning) greatly benefit from the current advances in AI, and in my estimation are already solvable with better harnesses with the current generation of models.
Those are essential for the powerful yet accurate electromagnetic motors that make these robots feasible. Oh, and batteries too. However, their supply is severely limited, even without China flexing its strangehold on the market. Last I checked there are no real alternatives or solutions on the horizon for the raw material shortage. Some estimates suggest the global supply currently is a small fraction (like 2% or even lower?) of that needed for mass deployment of the type Elon hallucinates about to pump TSLA stock.
The company that solves this or otherwise figures out a way to mass produce dexterous robots is going to be the next trillion dollar company.
Anyway, who can predict the future? There are still lots of things to do in IT in the age of AI - actually, more than before. The question is most whether you like writing software and doing planning and bug fixes all day for the next decades.
Secondly I think robot butlers are a bridge too far. The article explains pretty well why the spatial reasoning is problematic but so would things like battery life. And there'd be zero margin for error. The first time a robot steps on someone's cat or, god forbid, baby there will be a massive backlash. They have to be perfect. We accept LLMs hallucinating once in a while but a physical robot in your home cannot ever be wrong.
I think if we want to pilot autonomous robots it will be factory floors and other rote tasks. Maybe installing solar fields or something. Let them achieve a perfect safety record in someplace with more risk tolerance then take the next step.
I don't think it really addresses the article's concerns about long-horizon tasks or complex object manipulation. The objects were just bags and boxes, the manipulations were very simple, and each task was a few seconds long.
This seems likely to me, but I’m not a robot guy. We have certainly seen this 10x in value with 2-5x increase in context and horizon dynamic play out at least three times since 2019 with llms, and with self driving cars; having watched it I’m inclined to believe it can occur again.
Building factories and other robots is trivial. Just go and watch videos of any modern factory.
What's nigh impossible is doing complex interactions requiring finesse, "human touch", and at speed.
Even if you build a million robot-building robots, not a single one of them will be able to fold clothes, for example.
This is one of the dumbest ideas I've read here. Seriously; do you honestly believe, that we will be able to linearly scale all of the materials that go into robots?
You can't have a factory without motors, aluminum, steel, copper, quality control, CNC manufacturing, precision bearings, ball screws, harmonic drives, semiconductors, batteries, and skilled people to design, install, and maintain all of it. Every one of those is its own supply chain with its own bottleneck, and almost all of them are already constrained and at risk.
If anything, I would bet you $500 we will not be able to jump to the robotics age because we don't have enough skilled humans on this earth, due to the excitement about not training them because of their imminent replacement. That's before geopolitics - do you think China, if we actually built working humanoid robots, wouldn't immediately claim all supply for themselves and lock down the supply chain? Of course they would, it would be stupid not to. In the event of a working humanoid robot, every single dependency needs to be onshore, or it's not scaling.
This might hold true if the nation building the robots did not have a well established manufacturing base. However, I can see this working in China who has a strong manufacturing base where they can rapidly source all of theses components. It would make for a great experiment but you need working robots first. And that is where I think this idea falls short -- there are no capable robots to accomplish this self-replicating task.
With wheels you can't climb up a steep mountain. Hands and feet allow to hold tight.
With wheels you can't hold as much. Feet can be used as a replacement for an arm for some functions when arms are broken.
If a wheel is partially broken it's close to impossible to use. If a foot is broken I can walk on heel or toes only and bypass some issue.
It's not the best solution for many things, but the flexibility is a big win.
That flexibility isn't needed with machines, though. I can have many machines, I can replace machines, ... so I can optimize for the use case.
And sometimes the use case might be to simulate a human to trick instincts, which recognize "humans" subconciously. Be it in health care or be it an army of machine soldiers walking towards an enemy.
A solar field robot can only do one thing. A humanoid robot with a software model that allows it to install solar fields and then switch to a new model is a much more compelling use case. Eventually you're going to run out of solar fields to install, then you need a solar field repair bot, a solar field cleaning bot, etc. General purpose robots that can switch models to perform different tasks is not a dead end, but definitely a very hard problem.
No, these autonomous robots will be used by the military as soldiers. There will be no attempt to achieve a "perfect safety record". These soldiers are already here, pay attention to what is happening in Ukraine.
We have had remote controlled objects like this in various forms going back to WW2 and a few instances before. A wireless/fiber/starlink connection was added recently making them more robust. This is not a humanoid because they don't make sense from a power/weight carrying/so many other perspectives.
The only thing unique to the Ukrainian situation is the widescale usage.
Our shape is very flexible.
It's probably not the end-all most flexible format out there, but whatever format we end up with will probably be more similar to a human than to a 6df robotic arm mounted on a car. (It will probably have legs, for a start, not wheels.)
And as hard as physical things are, i assumed that the first robots would cost 7 figures they are already at 5.
It will not take 50 years to have a generic AGI style Robot walking around.
Similarly BYD in china has plants over an acre in size that run lights out with one human worker. My understanding is these are also vertically integrated, taking inspiration from Tesla and SpaceX.
I’m happy to learn more about the material world from people at any walk of life; right now we have humans building robots out of raw materials and robots building robots out of raw materials already in the world. My prediction is that humanoid robots building robots will scale rapidly in the next two years; please give me the bear case though, I’m genuinely interested.