Hacker News Reader

Comment by modeless | Hacker News Reader

Boston Dynamics is amazingly impressive as always. But I can't help but feel that they are a lot like the people doing computer vision with hand-written features back in the day. It seems inevitable to me that The Bitter Lesson is coming for their control theory approach. (http://www.incompleteideas.net/IncIdeas/BitterLesson.html)

Possibly the reason it hasn't happened yet is the inaccessibility of robot hardware; nobody else can try different approaches to compete because nobody else has anything like Atlas. Maybe that moat can keep them ahead for a long time, but if the hardware ever becomes cheap enough for real applications, it will be cheap enough for plenty of competition. They will have to adopt more ML methods or be disrupted. Of course maybe they already are, it's not like I actually know how their stuff works in detail, but consensus seems to be that it's still mostly control theory in there.

View on HN

Their methods are heavily based on hybrid planning, model-predictive-control, and trajectory optimization which have many parallels to popular ML methodologies [1].

Boston Dynamics problem is not "inaccessibility of robot hardware" or that "they have to adopt more ML methods"... There are more expensive machines in the world that are wildly profitable. Their problem is that it is very very hard to make a business case for humanoid or quadrupedal robots that is not better solved by more standard automation systems that (for example) don't have to be charged every 30 minutes or have to climb up and down the stairs.

[1] https://www.youtube.com/watch?v=EGABAx52GKI

> Their problem is that it is very very hard to make a business case for humanoid or quadrupedal robots that is not better solved by more standard automation systems that (for example) don't have to be charged every 30 minutes or have to climb up and down the stairs.

This reminds me of the realization that the difference between a Star Trek replicator and a microwaved can of Campbell's soup is about 30 seconds. We already have robots all around us. Microwaves cook our meals, washing machines do our laundry, dishwashers clean our dishes, etc. The time savings are immense but these devices do not resemble how humans perform the same tasks.

You make an interesting point but I think the difference between the replicator and the microwave is pretty great, as we recently learned from our worldwide supply train issues.

The can of soup had a whole factory, with robotic and human workers, management, metals, fossil fuels, storage containers, inspectors, transportation devices, highways, warehouses and retailers that it needed to get to your microwave. When any of those are bottlenecked there's a huge problem. It will likely forever be a fantasy but a Star Trek type replicator would streamline that process by a fair margin and have far fewer critical links.

A real-world example would be the difference between a old-fashioned printed newspaper with no digital distribution and an online only paper. The print version might seem superior in some respects but by the time it would get to McMurdo, the news would already be pretty old, vs up to the minute currency for a web browser. And once we start talking Shackleton or beyond, the difference only gets worse. In fact a good use case for an autonomous robotic presence is extraterrestrial exploration or labor. They could endure higher g acceleration, long periods in cold storage, don't need atmosphere, and don't experience existential terror -- as far as we know!

Important downside of replicator technology that it's not likely to be physically possible outside of TV. Otherwise I agree it's superior to microwaves (though curious about its techno-economics).

Why not? Perhaps we might not replicate everything but let’s say a food replicator might be very possible. No need to invent quantum whatever even.

At the speed shown on TV, thermodynamics is a problem.

Let's take water as an example. Say you've got a supply of all the atoms you need and a way to place them precisely enough to reliably get water rather than a mix of hydrogen peroxide and hydrogen gas.

When the atoms combine, they release as much energy as burning hydrogen in oxygen. Equivalent problems for virtually everything.

I'm not sure if it's provably impossible, but AFAIK there is no known way to cool this down fast enough to even be stable for the duration of the TV show's materialisation sequence.

That's why Picard orders "tea, Earl Grey, hot". The "cold" option is still on the roadmap.

https://xkcd.com/2570/

Tea. Earl Gray. Loud.

The energy doesn't have to turn to heat in the first place. A fuel cell can do hydrogen + oxygen = water + electricity.

Although I wasn't specifically thinking of those, catalysts are a great example of why I am somewhat sceptical of anyone who says "this thing is theoretically impossible" unless they actually prove it.

However, for the specific, fuel cells have a lot of waste heat. And even if they didn't, what you're describing in this situation would cause a mixture of ohmic heating and electrolysis in the water while it's being replicated — you need to get the energy out without the last 10% flash-boiling your tea into steam (or even ice cube or icecream sandwich or whatever).

To rephrase what I said before, it might not be impossible, but AFAIK nobody has a solution yet.

IIRC Star Trek style replicators and transporters get around that limitation by freezing time inside the assembly field.

I've never heard that lore before.

But we can't freeze time, and even if we could I don't see how that would help get rid of the energy of combustion either.

Oh that's easy, if you freeze time with quad-phasic, flux-capacitance, dilithium crystal matrices, then when they break down from the heat, the excess energy is dissipated into the reversed output of a toroidal sub-space containment field.

Essentially, a bit like a miniature black hole created on demand to dispose of present-day physics.

Easier to just microwave a can of soup, no?

But is it _cooler_ than freezing time to make your cup of chicken noodle?

The United Federation of "hold my beer, I got this": https://imgur.io/gallery/wpZ4w

I half-remember a point in Deep Space Nine where they started using the replicator or transporter to do surgery and I thought "ok, so you have a machine that is Dr. Manhattan, it can literally do anything. Where do you go from here?"

Turns out the answer is genocide (DS9) and screwing around with VR and time travel (Voyager). Pretty realistic.

And don't forget about the Heisenberg compensator!

How does it work?

It works very well, thank you.

( https://en.wikipedia.org/wiki/Michael_Okuda#Work_in_Star_Tre... )

I can either tell you or you can know, but not both

"a Star Trek type replicator would streamline that process by a fair margin and have far fewer critical links."

Would it? We're never really shown how replicators work or much else about the logistics of the Federation but I've always assumed the support machinery and network behind the replicators must be pretty enormous. Not only the manufacture and maintenance of them but just ensuring all the energy, fuel and raw elements are available and can get from A to B would be complex.

You’re absolutely right about the implementation details of Star Trek tech in the real world. But in the same way “The Lower Decks” is my favorite Star Trek series because we don’t know how much planning and logistics went into supplying the replicators on an Enterprise.

What’s the difference between storing massed of feed elements for replicators or storing canned goods? It’s really a matter of perception.

Once you eat the soup, your waste can be fed back into the replicator to be disassembled into constituent atoms and reassembled into more soup.

It would be very possible to live entirely isolated on a planet with nothing but a basic replicator and a way to power it. That's been explored a couple of times in the Trek universe

I have some news about fertilizer you might find disturbing.

> This reminds me of the realization that the difference between a Star Trek replicator and a microwaved can of Campbell's soup is about 30 seconds.

Well, 30 seconds plus the entirety of the farming, production, supply, transportation, and commercial infrastructure that gets that can of soup into your cabinet.

That said, I mostly agree with your broader point.

That and the replicator can make me a PS5 or a yo-yo to play with while my microwave heats up a can of soup.

The replicator could also create the microwave with the 30 seconds on the timer, with the soup inside, already running. Think of the time savings!

The five-minute hypothesis, also known as "Last Thursdayism", suggests that the entire universe, including our memories of earlier times, was created 5 minutes ago (or last Thursday).

So… have the replicator make the microwave with hot soup waiting on the inside, 0 seconds on the timer, ready to go "ding" the moment it's materialised. :)

Combined with the transporter, it can replicate directly into your stomach. Even more efficient!

That’s why there are no bathrooms

Now you are mixing up Star Trek and Harry Potter.

> This reminds me of the realization that the difference between a Star Trek replicator and a microwaved can of Campbell's soup is about 30 seconds.

Can a microwave make another microwave?

Self-replication is the killer feature of sci-fi replicators that people often miss.

We already to have robots everywhere, it's true, but they can't self replicate, so we still need humans in the loop. bipedal robots have the prospect of being the glue that combines all the other machines together, just like humans are.

Once that happens you'll have a novel machine that humans have as of yet never been able to make -- a Von Neumann probe.

I'm certain there are robots involved in manufacturing microwaves.

Yes but can a microwave make a microwave?

Not only could a humanoid robot make a microwave, it could even make a copy of itself.

And if it didn't have the parts lying around it could get into a car, drive out to the mountains and start mining the raw materials to build the parts.

Can a microwave do that?

I don't think microwaves or dishwashers can accurately be called robots. They're just regular old machines. A computer is not meaningfully involved with their operation.

You could certainly make an argument that a printer is a robot, though.

So are they robots if we throw an app frontend and some IoT thing at it so there's a computer involved?

I think that's just a microwave or dishwasher with a computer glued on. The computer needs to be actually integral to the operation of the appliance, and in general it needs to be involved with motion control.

So e.g. a self driving car could qualify as a robot.

I don't think OPs reference to them as robots is supposed to be taken literally, but rather in the sense that these are devices that substitute for many of the same human tasks that robots would.

> very very hard to make a business case for humanoid or quadrupedal robots that is not better solved by more standard automation systems

It's a general robot vs a specialized one. The only reason why it's a hard business case is usefulness vs price.

Why do you have a washing machine that sits there doing nothing 95% of the day? You literally dedicate usable space to this.

Obviously, the dream is iRobot (not the 2nd gen) and possibly like Startrek where working is optional and purely a pursuit of a dream instead of a necessity of living in the modern world.

> It's a general robot vs a specialized one. The only reason why it's a hard business case is usefulness vs price.

Yes. And as it is a common issue, we’ve seen it play out again and again: specialized wins for an incredibly long time, until the technology and production power leaps forward, and the generic solution becomes enough of a commodity to warrant using it in sub-par use cases.

I see the Raspberry Pi as the poster child of this: using a generic computing device to activate your garage door only makes sense because it’s so cheap to do so (at least while supply lasted)

I’d imagine humanoid robots will follow the same path: decades from now, as the cost to build one has ridiculously plummeted and we’ve already moved on to way more efficient solutions in specialized applications, we’ll have home humanoids tasked to do dumb things, because it will just be that cheap anyway.

I mean the thing with the raspberry pi in your example is there is even cheaper microcontrollers that are just about as easy to develop for as the raspberry pi. The ESP’s of the world are almost an order of magnitude cheaper and are specifically designed for such scenarios.

However those ESP chips ate the lunch of a whole line of other more specialized microcontrollers. So your argument still stands.

> The only reason why it's a hard business case is usefulness vs price.

This sums it up for me. If I could replace my dishwasher, washing machine, dryer with something that can do those things at an equivalent price/reliability then sure.

But a general robot requires far greater capability than any individual device and hence that is a hard ask.

But maybe you could lessen the general requirements by leveraging the specific machines.

IE if you can get it to use the washing machine, use the dryer, fold clothes and put them away (even if the wardrobe has to be a particular style/shape/format) - then "whole problem of washing clothes" would be solved.

And that would be popular. I'd renovate large chunks of a house if I never have to "do laundry" again.

This is SF talk, that has 0 to with realities of engineering, trade offs and cost.

Not sure if you mean Science Fiction or San Francisco. Either works.

which part the generalized humanoid robot? or the Star Trek?

> Why do you have a washing machine that sits there doing nothing 95% of the day? You literally dedicate usable space to this.

That's an idea. A mobile home robot that does not much more than pick up dirty clothes and dirty dishes and clean'em and file'em.

If we really needed that, all our homes would be already equipped with a system that move our dirty clothes to an external storage for a laundry service to pickup and a system that would pick up cleaned up and ironed clothes and would automatically store them in the correct wardrobe based on a tag system on the clothing. An automatic system similar to what manage big warehouses. It wouldn't be that complicated provided the house is build with that in mind and it could surely be retrofited to older houses by making holes in a few walls. It would be much more efficient than a mobile home robot.

The reason we don't do that is we are control freaks.

> The reason we don't do that is we are control freaks.

To be practical your service would need to pick up laundry at all hours of the day and be basically “on demand”.

Cats puke on your only comforter. Kids spilled something on your fancy jeans. You need towels, stat.

That being said, if the service did enough (say put the clothes away organized and folded and had a turnaround time of a few hours) people would probably find ways to work around the shortcomings of using a laundry service.

"Try our on-demand drone service that swoops in thru the window, cleans up that fresh cat barf, and then zooms away."

-CEO Catherine "Kat" Barfaway.

Without any profit incentives, we could probably have a few bigger washing machines per apartment house. Then it could have reasonably short cycles.

The business case problem is solely a capability problem. Flexibility is hugely valuable and would make a capable humanoid robot way more valuable than traditional automation in plenty of applications.

But Atlas just can't do anything useful without decent hands. And it seems to me like the mechanical problems and control problems for hands are actually much harder than for legs. Boston Dynamics doesn't have much going on there yet, at least in public.

I agree on both counts. The ability to repurpose the robot as needed for a virtually infinite range of tasks is worth a lot…

As for hands, I wouldn't be surprised if they're working on something in that realm, but good hands alone aren't worth a whole lot without the rest of the robot and they don't make for particularly snazzy demo videos, so they haven't been shown off (if they exist). Now capable hands on a capable robot on the other hand… that's a recipe for a viral video.

meanwhile thirteen years ago, high speed robotic hand:

https://www.youtube.com/watch?v=-KxjVlaLBmk

> problem is that it is very very hard to make a business c

Not sure if you can apply the "business" label to it but I'm pretty sure that the US Military is quite happy to throw money at this, no questions asked. Also strange how the guy at end was talking about possible use-cases mentioned stuff like construction but not the battle-fields or suppressing dissent.

I look at these robots and see the business case that they could be used to take care of the elderly. They could literally do the heavy lifting in care homes, leaving caregivers to do the more gentle tasks.

There’s a big market there that would only grow, given current demographics.

I don’t see the Tesla bot catching up with Boston Dynamics any time soon.

Can't most of the "heavy lifting" be done by more specialized machine anyway? I mean besides electric beds and chairs there isn't much left. I'd rather give powered exosqueletons to caregivers if they need more power to lift an overweight person than use a robot for that application anyway.

Maybe, but based on what I saw in my mother and grandmother in their old age, I think androids in general (not just Boston Dynamic) have a long way to go just on the "accidentally terrifying" front, let alone capabilities. Even for me, things like the following video are the wrong side of uncanny valley:

https://youtu.be/uXlQuTRSmzc?t=270

The business case will eventually be selling robotic companions to chronically lonely people.

Loneliness is arguably the biggest under-addressed problem in the wealthy developed world.

I can kind of see that, in terms of robot pets or sex robots. Clearly robot pets are going to be more marketable. Sex bots will definitely be a weird niche, until some very revolutionary (if not scary) breakthroughs happen.

Boston Dynamics seems to be making a much broader business case than just lonely people. For instance warehouse workers, factory workers, bartenders, dangerous jobs, interplanetary explorer, etc... There are a lot places where such humanoid robots can end up, which is way beyond a "lonely" people market.

I feel exosqueletons are a more interesting approach anyway. Maybe robots would do less errors but I don't know, most of the times I tried an AI it would understand most of what I asked but if I would miss to precise some details but it would still do the job while an human would ask for more precision on the request.

I would hate my robot to break stuff or put stuff in shelve C while I wanted it in shelve A and have to ask him again. Sure sometimes human workers also assume things instead of asking for more precise direction but I have the feeling we are more easily puzzled while an AI seems to be trained to be overconfident in order to be useful.

Also wage supression towards zero. Some companies do more for UBI then the biggest revolutionary ever could do..

A follow-me feature that gets 500lbs of shingles up a ladder and onto a roof would be useful. Granted, probably way more expensive than human labor.

Make me think does living things' evolution make of pure functional utility purpose?

This was my thinking. The robot in this video replaced a dumb waiter.

With some work, it could probably even replace a smart one.

I disagree with The Bitter Lesson, or at least one particular example that the essay gives: chess programming.

Modern engines would crush Deep Blue with fewer resources and even without NNUE, and this is because of specific search techniques and search improvements that we've discovered. For example: formulas that work well for late-move reduction; continuation history tables and the update formulas that work well for them; internal iterative reductions; refinements to futility pruning, reverse futility pruning, delta pruning, multi-probcut, and so on. None of these were found by meta methods and none of these emerge from scaling computation. But it's also the case that none of them resemble anything that a human does when they play chess.

So I don't think I agree with the first point of the bitter lesson, that general methods beat specific methods, and that scaling beats efficiency. But I do agree with the second point, that performant systems often behave in ways that don't resemble human thinking, and we're often hopeless to personally understand performant systems' complexity.

Chess engines aren't a done deal. Besides NNUE (which I would consider more general) that are outright better than handcrafted eval functions, the dust hasn't settled yet between the more specific AB pruning stockfish vs the more general MCS lc0.

Yes, stockfish has been generally slightly stronger, but in the near future with better NN architectures I could easily see this changing.

And even then, chess is probably an example of something very narrow that you can deal with specific methods. Humanoid robots are probably the opposite, and you probably need AGI for them.

I'm not an expert in this, so question... Doesn't crush the AI AlphaZero the software Stockfish? Yes, AlphaZero needs more resources - especially for training. But it learns chess from scratch by self playing - in a few hours, without any data from over 1000 years of humankind chess experience. Deep Blue was more like Stockfish than AlphaZero. They called it AI - but its not an AI in the sense of AlphaZero, it was an expert system like Stockfish with special search algorithms. They just needed such a big machine, because this is just like CPU power developed. So I think the term "Bitter Lesson" is a very good match for chess too. A modern AI plays 4 hours from level "very stup... beginner" to "better than any human or machine in the world". It doesn't need a bunch of human super experts, that inject the complete human chess knowledge into specialized databases. This is bitter - isn't it?

>Modern engines would crush Deep Blue with fewer resources and even without NNUE

The point of the essay is that modern engines with NNUE beat modern engines without NNUE. So all those human coded features that used to be used to evaluate positions are no longer necessary.

I watched (and even more so heard) his great, great, grandfather hop around in Raibert's lab next door to ours at CMU in 1983. A fellow freshman who worked with him came in to the lab one day with a four inch tall wind up carnival toy that does flips. He said to us that their upgraded $10m pogo stick would soon do flips - and he was correct. Of course my thought at the time was "how do they get funding for this nonsense?" But to see how far Atlas has come in these 40 years is just fucking mind blowing.

Video for context: https://youtu.be/XFXj81mvInc

Thanks

I can't speak for right now, as I last worked on this stuff 8 or so years ago, but I briefly contributed to a DARPA robotics challenge team and simply working in ROS as a whole is an immensely difficult process to really get anything to happen with atlas robots. It's not just something a small number of engineers can quickly ramp up on and be immensely productive immediately, you'd need a pretty big team working on stuff over a period of a couple of years to really take things to the next level.

There was a recent project that combined transformer breakthroughs (such as those used in ChatGPT) with human motion.

[1] https://guytevet.github.io/mdm-page/

[2] https://news.ycombinator.com/item?id=33029522

If anyone from Boston Dynamics is reading this, please explore this line of research :)

Would be awesome to command Atlas through plain spoken language

I bet everyone in robotics is thinking about combining large language models with robot control right now. I expect some pretty awesome demos in the next few years. Robots navigating real spaces and performing actual mundane tasks from natural language commands feels a lot closer than it did a year ago.

Such demos already exist. I watched one last year.

The trick is a (by now) common one: you train the LLM on a simple programming language that's interpreted by the robot's software. So natural language goes in and the LLM then thinks out loud to figure out the needed steps, then emits Logo-like instructions to control where the robot goes. More conventional robotics AI then takes over to handle movement of the machine, the arms, etc.

I wonder what NSA and the like are doing with LLMs in context of cryptography.

Is there a Shannon information theoretic limit that would make that thought go away? (Doesn't have to be end to end. Could just hugely reduce the search spaces..)

-- ps

2008:

https://www.cs.unc.edu/~fabian/papers/oakland08.pdf

> nobody else can try different approaches to compete because nobody else has anything like Atlas.

But but you know why they have Atlas? They went to a garage and clomped on metal and made it. It is not like some control theorist wandered into the woods and found the right robot. They were building a series of simpler ones, trying their ideas and seeing what works and what doesn’t.

The "back in the day" hand-written features are still the to go method in safety critical applications and aerospace (autonomous landing etc), excluding of course self driving cars.

Sometimes you just cannot replace a provable method with a black box neural net, even if the latter is in fashion.

That Bitter Lesson link is an interesting counterpoint to Stephen Wolfram's recent post on WolframAlpha|ChatGPT. Can the world be learned through statistical methods? Clearly it must, as humans did so as a collective. To calculate the distance from Chicago to Tokyo, in the Wolfram approach at one time someone had uploaded Earth_ellipsoid_model and the GPS coordinates of all the world's major cities into Wolfram Alpha. Is doing so falling afoul of the Bitter Lesson?

that's the thing i learned in my robotics class, with a from-scratch robot project. the mechanics are the hardest thing. the electronics the 2nd hardest.

And then after you've solved those problems and built some robots you have 2-3 years of software and debugging before anything works reasonably well.

One problem I see in this particular domain is autonomy condition. We can't put that much computing power on board of Atlas/Spot (also minding literally power consumption) and should not make them depend on external computation, because it greatly reduce their application scenarios.

what makes you think they are explicitly choreographing each facet of movement?

they show they run simulations first, so they first explore a large range of ways to achieve goals, that would be incredibly expensive to manually micromanage.

obviously they "compile" the highlevel tasks into lower level "bare metal machine language" ones, to make a metaphor.

the people who design a domain specific "compiler" need "intermediate languages" for interpretability and debugging.

the fact they give us a peak at their "debugging" capabilities should not be misinterpreted as a bunch of engineers hardcoding in the "assembly language" of robot control.

it seems bizarre to view the advanced state of their software tool suite (inspectability, visualizations, ...) as a sign of primitiveness.

The subjective experience of becoming aware of the full frontal details of a car when popping open the hood doesn't suddenly demonstrate the "primitiveness" of the car, but rather the subject's prior lack of awareness of how things work...

Reply on news.ycombinator.com