Hacker News Reader

Comment by ACCount37 | Hacker News Reader

"The vision" is that instead of building 9999 specialist robots for 9999 different tasks, you mass produce one robot model that can do all of them.

Less efficiently, sure, but for the manufacturing, logistics, maintenance? The economies of scale are immense.

The reason why we weren't doing exactly that back in the 80s isn't that universal humanoid robots somehow weren't desirable. It's that for a universal humanoid hardware to be useful, you need a fairly universal AI to back it.

That "universal AI" was nowhere to be seen back in the 80s, or the 90s, or the 00s. Now, we finally have a good idea of how to build the kind of AI required for it.

View on HN

> you mass produce one robot model that can do all of them.

This just occurred to me: do standard industry robotic arms not fit that description perfectly? They're not specialized for any particular task, the only customizable parameters are the size and the end effector.

They can move around car bodies or seats, or pick up an airbrush. They can probably be installed with a five-fingered hand, or onto a giant human torso, should such tools somehow made sense for some applications. They feel like the generalist robot that meets most of the expectations for the hypothetical factory humanoids, sans being a humanoid. I mean, I get it, but aren't those existing bots just what "the vision" calls for?

I can see that for the factory floor, but there's no particular reason for "it" to be "humanoid".

It's basically a robot arm with mobility at that point, and if you need more than one, just have more than one robot wheel into place. There's no particular reason to have two arms.. one, or three, or five are all sensible numbers. Heck, a chassis supporting a variable number of arms and other appendages (sensors and so on) is plausible, and the control system looks more like an ant-colony mind than a human one.

Which is a long-winded way of saying, there's no particular reason to link embodiment and cognition at the individual arm level in a factory scenario.

On the factory floor, all the tasks that were a good fit for "a robot arm bolted down to the floor next to the assembly line" are already performed by robot arms bolted down to the floor next to the assembly line.

What remains is all the weird and awkward automation-resistant tasks where "just get a human to do it" is still easier and cheaper than redesigning everything to maybe get old school automation to handle them.

This is the kind of niche humanoid robots are currently aiming at. It's no coincidence that at least 3 companies trying to develop humanoid robots have ties to vehicle manufacturers.

Makes sense - but I'm still not clear on why it would need to be humanoid in form-factor or indeed self-contained/autonomous and not a more flexible, mobile, multipurpose robot-arm-like-thing?

Like is there any particular reason for it to be about 6' tall with exactly two 3' long, three-jointed arms rather than any of the other possible permutations for those things?

You want robots that can do tasks humans can do but robot arms can't - which has a way of driving the design decisions. The more you move towards "flexible, mobile, multipurpose", the more you move towards the humanoid frames.

Just look up how vehicle interior assembly is performed now. Look at all the things that are still done by humans - all the different assembly stations, all the loading and unloading, all the installation operations, all the panels and wiring harnesses and plugs and bolts.

Then try to come up with a robot frame that would do all of it - every single operation that's currently done by a human - while being significantly less complex than a humanoid frame.

The design space constraints would choke the life out of you.

But then you need one robot that's got 9,999 specializations. A human can't actually do 10,000 things. You need 10,000 humans, each that goes through a specialization process (training, developing muscle memory, etc) that builds both physiology and mental capability specific to the skill. Not only does the robot need to be capable of every incredibly difficult physical skill we learn, it needs a matching program.

It's impossible to do this in a general way. This could theoretically be scalable (produce the robot and have 10,000 companies all develop their own specialization routines), but the hardware (both the parts as well as neural interface) needs to be as capable as a human body, which isn't even remotely true. The physical robot will always limit what skills it can learn, on top of the difficulty of programming the skill.

I think we're hundreds of years away from making a robot that's as capable as a human. We would get there faster with synthetics or cyborgs. Create a human body without a brain, use Neuralink to operate it. Until then, specialized robots are the only thing that will scale to 10,000 skills.

The words you're looking for are: "transfer learning".

Currently, dedicated robotics datasets are pathetic - in both the raw size and domain diversity - compared to what we have for generative AIs in domains like text, sound, video or images. So adding any more data helps a lot.

If you trained a robot to fully strip down a specific e-scooter model - whether for repair, remanufacturing or recycling - that training data would then help with any similar tasks. As well as a variety of seemingly unrelated tasks that also require manual dexterity, manipulation and spatial reasoning.

Those "9999 specializations" all overlap in obvious and subtle ways - and they feed little bits of skills and adaptations to each other. Which is why a lot of the robotic companies are itching to start pushing the units out there as soon as they are able to perform some useful tasks. They want that real world training data.

The question is how much less efficiently. The point on which it all hinges is that it needs to be a little less efficiently, but the reality is probably that it's so much less that it's no longer really viable.

Like, Digit costs a quarter mil and is rated for 10 thousand hours. It can stack boxes. For that price you can turn every box in your warehouse into an AGV and they'll last you forever.

The other part of this vision is that out of 9999 different tasks, only 1-2 are routine enough and valuable enough to merit a purpose-built robot. But a generalist robot could theoretically do an almost infinite variety of low-stakes tasks that people currently do.

The other part of that argument that I've seen is that they'd be able to use generic tools made for humans.

Reply on news.ycombinator.com