Robot Learns to See in 30 Minutes (2022) | Hacker News Reader

Robot Learns to See in 30 Minutes (2022)

antonilo.github.io

229 pointsstoniejohnson93 comments

Via simulation + meta-learning they can make a blind robot walk over arbitrary terrains solely via proprioception.

They add a visual component, which with 30 minutes of real world data, through a technique called Cross-Modal Supervision (CMS), is able to improve walking performance significantly!

They can even distort its vision (turn the camera) and it learns to adapt!

Paper: https://arxiv.org/abs/2211.03785

Advisor Presentation: https://www.youtube.com/watch?v=f6fDpKDxpL0

Note: This is not my work; I originally found the presentation above.

When one robot learns, millions of them learn

When millions of cars learn, they all can send data back to the mothership which crunches the numbers and they all improve

Even without AI, this is how technology has always advanced. Humans simply improved things and copied it to all the implementations.

So basically, when it comes to competing with robots or software, it is a losing proposition in the long run. No biological animal can evolve faster than this.

And yet our politicians ban human cloning and designer babies. Along with climate change, it’s like our politicians want human extinction.

The same hubris behind "let's clone humans" brought climate change. And the development of AGI is a similar self-destruct bet. The extinction will be premised on those things being allowed, not on them being banned.

building on your thinking... what if humans are merely a point on an unfurling of intelligence/hacking, which will eventually reach its natural end, as all things do? just like the universe will as well; it couldn't foresee how its processes will yield its end

We don't exactly know the nature of 'The Great Filter'.

But maybe we'll find out.

Sure, but what about it?

As the member of a species being extinct, I wouldn't like it and just "roll with it". Same way a race being told that they're there to serve, and if needed, be extinct to bring forward some other entity wouldn't.

Until we reach that "natural end" we shouldn't go "gently into that good night".

> we shouldn't go "gently into that good night

concur. the challenge in front of us (and always has been) to surf the waves we can find in this universe

One of me is quite enough.

It's not you that I'm particularly worried about. It's Bill Gates VII and Elon Musk XIV that I'm worried about.

Not the worst people to have more of.

I strongly disagree with that.

Oh, so you would rather clone Hitler and Stalin than Musk and Gates?

Why not neither?

That's got to be the mother of all false dichotomies.

Godwin aside, the point is valid. When I look at Musk? Sure, he's overrated, acts like a nerdy teenager who never grew up — "Rockets are cool! Sports cars are cool! I know, I'll put a sports car on a rocket!" — but even where I think he's surrounded himself with sycophants (IMO Boring Company, Neuralink, based purely on what I'm told by civil engineers and brain scientists) not really evil.

Misaligned and powerful, so an easy example of how an AI can go wrong; but you don't have to reach for Auschwitz or the Holodomor to find people much worse to clone than him.

As for Gates? The worst true thing I've heard about him was Internet Explorer. And possibly Clippy.

Maybe I misinterpreted your “I strongly disagree” comment. I thought you meant Musk and Gates are the worst people to be cloned.

Aren't you embarrassed to field such peurile arguments? Really, that's an insult to everyone else's intelligence.

Ok, I'll invoke Godwin and end this comment subthread.

I don't lose much sleep over it. They'd be too busy plotting the untimely demise of their other clones.

If you had a bioidentical braindead clone (ie. from a body farm), you could harvest it for new parts. You could replace failing organs, rejuvenate youth, and (with the technology to do so) perform a full head transplant to rid yourself of all cancers, cardiovascular damage, pulmonary damage, and other age-related injuries. You get a fresh immune system to boot.

Because it's bio-identical (ABO/Rh/HLA), no immunosuppressant is necessary and you can continue about a normal life.

Regenerative cloning should absolutely be a thing.

Well, brain and face diseases and cancers would still be a problem. "rejuvenate youth" is a magic phrase in there that promises a lot but doesn't actually mean much.

> "rejuvenate youth" is a magic phrase in there that promises a lot but doesn't actually mean much.

I disagree; though the phrase and its synonyms hides a lot of complexity — and I have to temper my optimism about research with the knowledge of millennia of snake oil — we'd recognise it if we saw it. (Unlike, say, "is this AI/animal conscious?", where we can't agree what the question means).

Would we? What all comes with "rejuvenate youth"? We have technology right now that makes one LOOK younger. Is it more than that? Where does it stop? Do our brains have to reverse their development so that the bits that don't fully congeal into our late 20's/early 30's are removed? Do we need to become impulsive, awkward, hormonal weirdos again? What's youth? 35? 25? 15? Does youth requiring scraping out all the plaque out of our arteries? Do we remove our scars? Do we get acne again?

A lot of what makes us us with all the wisdom of our age is also the bits that lead to the eventual breakdown and decay of our body. I think we'd need to lose ourselves to get our "youth" back. We're hoping for some magic where we keep all the good with none of the bad.

We've pretty much already done this with rats. The pop-sci summaries of the research I've seen (with all the caveats implied) say this improves cognition and memory, as well as general physical health and (because rats) fur colouration and quality.

Sadly, the usual problem is that rats aren't really very close to humans, merely closer than the other stuff we research with.

I assume none of that involves creating a braindead clone of the mouse?

Not that I heard, though it doesn't matter either way with regard to the specific issue I was critiquing.

Yeah, fair enough. The goal isn't necessarily magic. But I feel like the way to get there, at the moment, mostly is, especially with something like comparing a mouses lifespan against ours.

Sounds awful. Like my life of course and would be bummed if I fell over and died due to organ loss but OTOH I don't share the fondness for immortality or crippling fear of death that seems to affect many technologists. It seems just as natural that things should have ends as beginnings.

I don't fear death, but what I don't like are constant health problems when you get older. My grandma complained about her health problems constantly - it was so annoying.

There is also a strong influence of mindset here: there are people who are psychologically able to handle far more painful situation and still enjoy whatever satisfaction reachable to them, when other will lament all day along how sad they feel when they have best health and great social situation but can’t afford whatever luxury thing that might be above their current wallet.

To be clear, this is not to blame your grandparent. We all deal with situation with behavior that life forged into us. But if your opinion is that free will exists, then you should probably agree that there are ways to improve ourself.

Old people have measurably more health problems though, it does look uncomfortable. If there is an option to avoid that, it would be better.

Well, start take care of oneself, regular but not exaggerated physical exercices, good nutritional habits, sleeping enough each day and having balanced social interactions: these are all well known and documented.

Not ways to live to 200.

No cloning can produce what your body accumulates while interacting with enviroment. Lookup epygenetics.

Surely that is a good thing in this case?

Borgdom awaits

No biological animal can evolve faster than this.

What’s the point of evolution for something which isn’t alive and has no will to live and no environment that it really needs to be part of for survival?

I think about this when we talk about bots that continually improve? Continually improve to where ? To what end ?

I’m asking out of curiosity not out of malice btw, it’s just an interesting thing to think about.

I’m hoping that we eventually realise that life is pretty awesome albeit slower and learn that even a bird is really an amazing thing.

There is no point to evolution in the technical sense of the word. Evolution is just a dynamic. Dynamics don't seek purpose.

The thing that makes evolution a special dynamic is that everything that exists and you get to perceive is either the result of evolution, or compatible with it.

Will to life is effectively an artifact of evolution. Things with ways to express some form of persistence and reproduction have an implicit will to live.

Leaving aside whatever goals AIs may be programmed for, if at some point a dynamic for persistence and reproduction appears it will dominate any competing systems that don't have it on that basis alone. Systems that incorporate and propagate useful features for their survival will have capabilities to adapt and persist that are very limited in known lifeforms.

This is assuming that evolving at 1 million times faster than other life is actually possible or advantageous. There may actually be a reason why faster != better.

Of course there is a point to evolution, which is to utilise the environment to preserve your genes and pass them onto the next generation.

Evolve to slow though and you die, evolve too fast, you’re no longer compatible with your environment and also die.

Bacteria can reproduce in about 30 minutes, I think. That's in the order of 1,000,000 times faster than a human.

As I'm not a biologist, I would only be guessing if that corresponds to evolving 1 million times faster, or if the relative complexity of the organisms makes it either more or less than that.

if "evolving too fast" worked out worse for survival than "evolving more slowly" then it wasn't evolving at all

evolution always wins, by definition - whatever wins, that was evolution

> Of course there is a point to evolution, which is to utilise the environment to preserve your genes and pass them onto the next generation.

that is just a concrete mechanism to evolution, albeit a special one for humans as we're a part of it

Some would even say that birds are smarter than humans, in fact.

> What’s the point of evolution

Evolution by natural selection has no objectives, goals, direction or, for that matter, point. It has no purpose, it just happens. Every living organism on earth is the result of an unimaginably large series of random mutations across an equally unimaginable span of time, all completely absent of guidance.

AI, however, is different. We set the objectives. We decide where we want it to go. And, based on this, we setup artificial selection to make it happen.

A suitable cost function to optimize. It could be moving faster, reducing energy consumption. Or maximizing the harvest of the use case type called "money" from the environment called "humans". Whatever that means.

> No biological animal can evolve faster than this.

Are you sure? The robot is limited by its gears. However, animals have cells that can adapt and change through a super sophisticated mechanism called DNA.

We are still way far from reaching living organisms complexity. All stuff we make is very fragile compared to that. Once we are able to modify real humans, you'll know we have the real deal.

> The robot is limited by its gears.

Wait till they hook up the robots to other robots including 3D printers.

At which point does self-assembly and 3d printing overlap?

Is there some minimal viable size for a 3d printer to be shrunk down to, aka, cellular level?

Is there diminishing returns or does one eventually just go to nanites for this?

> Is there some minimal viable size for a 3d printer to be shrunk down to, aka, cellular level?

Great, not only chemicals, bacteria/virus, nuclear radiation, drones with guns can kill us all very efficiently, bot soon nanobots can devour us alive.

In principle you can pick-and-place individual atoms with an electron microscope. But if you think macro-scale 3D printing is slow, you'll be flabbergasted by how much slower this is.

Also Bios have millions of years of pre-trained reservoir DNA to put into use as/when needed.

What is preventing robots to incorporate all that information in their own models?

Its all undocumented legacy code?

I'd say that usually holds us back more than anything else.

Pretty much 99.99% of crime is related to our evolutionary process, imo. The edge cases of mentally ill people etc are low, but otherwise racist/homophobic/theft crimes are all just tribalism/being afraid of the other/resource contention at the end of the day; society and its rules are just a patch on top of millennia of forming small groups and marching over to hit other neighbouring small groups with rocks.

Sentient life would be both impressive & scary in that it can theoretically optimise itself far, far beyond what we can. Then again it could go the other way too: aggressive nanobots that start disassembling humans getting into tribal-like wars with other groups of nanobots to fight over resources aka humans to melt down.

> No biological animal can evolve faster than this.

Sure one does. Us.

When one human learns something, they communicate about it and often, they all do. As always, robotics and automation doesn't enable any fundimental new capacity other than what required many humans to do, now requires less humans to do. This just gives more control to the humans who own the robots.

The AGI apocalypse is just the fear that capitalist entities and morales will spread and grow faster than humans can compete with. AI already exists to further their reach. We already have profit maximizing paperclip runaway super intelligence. We call them tech companies, and they flounder occasionally but they learn and adapt efficiently enough.

What seems clever to me is the super cool hack of getting sensor data while the video data is captured, then training a model on realtime video and backward time-shifted sensor data; result is a network that can predict sensor data 30 seconds ahead of time with visual data. This is then used to feed the walking model. That's the sort of super 'dumb' thing ML folks can do now; that task would have taken many many person years of work before, now, I speculate thinking of the idea to testing a network would be under a week. Verry cool.

I often think of this as the corollary to the bitter lesson: it is enough just to break some particular relationship between two things, and then throw compute at it. In the process of trying to recover that relationship, the model learns about the underlying real relationship - sometimes it even learns about each thing itself.

> training a model on realtime video and backward time-shifted sensor data

Do you mean forward time shifted sensor data? I.e. using sensor data 30 seconds from the future as a label, and real time video as input?

I think we're saying the same thing. Training pair at t0 is video_t0, sensor_t30. So we have shifted sensor data back 30 seconds, or put your way, we look to sensor data 30 seconds in the future.

Oh yeah you're right, thank you. I think the way you phrased it makes the most sense now.

Meanwhile, it takes the human brain months to adapt to visual input after blindness: "This research suggests that the brain is able to adapt to getting visual input, even after 10 years of blindness. That said, it takes a little while before this ability to really process the images kicks in. Right after the second surgery, patients are getting input to the brain from their retinas, but their visual memory is still poor. After about a month, though, the brain has found ways to process this visual information, and, a year later, their brains do just about as well as those of people who had sight from birth." Source: https://psychologytoday.com/us/blog/ulterior-motives/202210/....

Another good example: apparently it took 10 days of wearing upside-down goggles for a researcher to see things normally.

https://en.wikipedia.org/wiki/Upside_down_goggles

https://www.theguardian.com/education/2012/nov/12/improbable...

I liked the example in the article of teachers learning to read documents upside down (presumably because they do this a lot when students are in front of them).

> "After about a month, though, the brain has found ways to process this visual information, and, a year later, their brains do just about as well as those of people who had sight from birth."

Hey, that's still pretty impressive! The brain starting to receive a signal that it hasn't before and still being able to eventually adapt and figure it out is wonderful!

“After training our robot on a large set of terrains in the Berkeley campus, we verified generalization of the visual policy in the Stanford campus.”

Sublime. Almost a koan for meditating on AI research.

We did something similar but much simpler in 2009 where the robot learned where its sensors where and which way its motors moved it, and how to use them.

https://www.youtube.com/watch?v=kk-H--e6PUs

Last author, J Malik, has made many important contributions to computer vision. One I'm very grateful for is spectral manifolds.

this is actually an understatement - a nickname for him around here is The Mighty Malik

He recently gave a talk that covers robotics for a general audience: https://www.youtube.com/watch?v=f6fDpKDxpL0

It's not the learning that's the magic. It can be 30 minutes or 30 days. It hardly matters. Where the sparks fly is that learning can he transfered in .3 seconds to 300,000 other robots. And then updated as needed.

Now compare the to humans.

The fact that it is learning online in fractions of a second to prevent falling (I would say) is pretty magic!

Or you can train multiple ones with a kind of evolutionary approach - e.g. you train 100 at the same time and after a day you take the one with the best results, overwrite the rest with it's weights/config/etc. and repeat until satisfied.

Right. But it's not the learning. It's the ability to scale what's learned.

Humans learns one by one by one effectively in silos.

Robots? There's no comparison.

It's that learning is local and quick that makes it able to be fully general. New environment? It can learn that.

Can’t humans do this with…the internet ?

Learn to walk?

I mean who knows with a computer / brain interface what we’ll be doing?

You do realize there are already plenty of species on Earth that come out ready to walk, forage, swim etc?

Over the internet, I can transfer the knowledge of Spanish to you in around ten years. In person it will be more like two.

Won’t be long until we have hardware that can change this no?

You won’t understand Spanish but it will translate in real-time for you ? Just like how this robodog doesn’t know what a stair is but knows how to navigate it. It doesn’t understand the way you do but it’s still useful.

Humans not only "learn to see". They learn sensor fusion, which is the aggregation of your senses.

Read about vertigo and the ear, mirror box therapy, etc.

I can't find it in the paper - is the inference ran on the robot directly or remotely on a more powerful computer?

Likely local: "We build a custom model to predict the future to allow inference on limited compute."

I think the RealSense line has been discontinued, but a different company called Luxonis has kind of taken its place using the same Intel/Movidius platform.

realsense still exists. Please google stuff before dropping misinfo in the comments.

>This work was supported by the DARPA Machine Common Sense program

Coming soon to a neighbourhood near you!

Looking forward to seeing the code! Any sense how soon “soon” might be?

ok, which jobs is this one going to steal now?

Puts pack mules out to pasture. (Non-trivial application.)

Dog walker.

Depends on whether or not the right people have seen the right Black Mirror episode.

dogs

Guards for human cattle

Reply on news.ycombinator.com