One word of the title is not obvious. Here's how it is defined in the text:
> Ephemeralization, a term coined by R. Buckminster Fuller in 1938, is the ability of technological advancement to do "more and more with less and less until eventually you can do everything with nothing,"
> Consistent with this theme, I believe the solution to scaling up generalist robotics is to push as much of the iteration loop into software as possible, so that the researcher is freed from the sheer slowness of having to iterate in the real world.
...
> The most obvious way to ephemeralize robot learning in software is to make simulations that resemble reality as closely as possible.
Yeah, I must admit that from the title I thought this was going to be an argument that we should factory reset all robots daily just to ensure they never become conscious, or start drifting from their default settings in other ways. "Mindwipe them all, it's the only way to be sure."
I think it lets me know if the article is worth my time. I usually glance at the comment sections first because small titles often aren't enough to understand what the article is about.
Supporting this, I would say I'm reading the first couple comments on a submission far more often than reading the submission itself. It's not even close, probably three or four times more.
That's another signal, yes. Lots of things are interesting or important to the HN community that I don't care about personally, and even if that weren't true I don't set aside enough time to read the whole front page every day.
We used to say that AI development should be "embodied," because only the real world was complex and structured enough to train up an intelligence. Funny how the tables have turned!
My experience matches the author's exactly. You must check your robot against the real world to ensure it works, but simulation is the only feasible way to train and evaluate your system at scale.
The challenge is that simulators are kinda sorta almost good enough, but no actually they're not, and you need to solve a bunch of problems around the sim2real gap. The problems range from doable (colors look different under certain lighting conditions), to hard (simulating how other robots and people react to you), to still impossible (the feeling when a USB C plug has snapped into place). Each day we close the gap a little bit, but there is still a ways to go.
I still think that AI development should be embodied, and am willing to entertain the possibility that you don't need simulation at scale to train it (e.g. some clever unsupervised learning algorithm).
Evaluating a system that can do millions of things is simply impossible in reality, and that alone necessitates some kind of software-based evaluation metric. I agree that there is a ways to go - but I think trying to make simulation more like reality is a safer bet than having a low ceiling on iteration speed in real.
I think it's the siren song of thinking there were going to be more ways for a practical AI to cheat, and it turns out all the corner cases actually matter. Everyone had their fingers crossed that 'the world' was mostly accidental complexity instead of intrinsic. And to be fair I'm not sure that you historically had a choice in this. If you were doing AI research in the 80's or 90's there just weren't enough compute cycles available to model anything complex. Either you believed the gap was narrow and you found another field to work in, or you hoped that it was wide and that you were going to discover just how wide it was.
Every mammal has a model of the universe in their head that they work against. Mirror neurons expand that to modelling other mammals. I don't know how you can reason about an environment if you don't have a useful model of it first and foremost.
Having worked in robotics, I can absolutely vouch for the usefulness of simulation. It's not a 100% solution by any stretch, but it will absolutely help you smoke out the stupid, low hanging bugs, freeing you to concentrate on the messy shit that happens in the real world. When I was in the field, simulated tests were a critical part of our CI infrastructure, and warnings would be sent out if something failed in simulation. I even jokingly proposed a coat of arms for our software team featuring an "AUV naiant" blazon and the motto "Fungitur Simulatoris" -- dog latin for "it works in the simulator" (literally "he/she/it performs for the liar").
I've known adventures, seen places you people will never see, I've been Offworld and back… frontiers! I've stood on the back deck of a blinker bound for the Plutition Camps with sweat in my eyes watching stars fight on the shoulder of Orion... I’ve felt wind in my hair, riding test boats off the black galaxies and seen an attack fleet burn like a match and disappear. I've seen it, felt it..
I've seen things you people wouldn't believe. Attack ships on fire off the shoulder of Orion. I watched C-beams glitter in the dark near the Tannhäuser Gate. All those moments will be lost in time, like tears in rain. Time to die.
I find it disconcerting that these two quotes allude to attack ships, or warfare of some sort. Of all the interesting things that were seen that could've been mentioned, and war was included.
The character saying those things was designed with a 4-year lifespan to do nothing but fight in wars. This is his moment of character development: he was a tool of warfare, but he found beauty in those moments, proving he was capable of something else besides killing.
It's a major point of the movie.
Note that in a sense, it mirrors many developments in automation in the real world. Military drones and whatnot.
They're different versions of the same quote. One was the original version in the script and the other was what Rutger Hauer actually said in the movie.
The memories are about war because the android speaking them was a combat model, so his entire life had been spent fighting in wars.
Blade Runner is an excellent example of the power of collaboration. Hampton Fancher's script Ridley Scott's direction, Vangellis, Mobius, Syd Mead. Rutger Hauer developed his character more in ways that are famous, and Edward James Olmos invented cityspeak.
"...but the writing is on the wall: software is coming for hardware, and this trend will only accelerate."
I wonder if this is perhaps one of those cyclical things. Software comes for hardware. Then we all remember how much good hardware simplifies software and multiplies capability, so hardware becomes very important again, etc.
What the author meant here is software is coming for mechanical hardware, like gears and motors. I.e., with more advanced software you will be able to use cheaper gears and motors. As far as electronic hardware, forget about it. You will need more and more of it. His vision of replacing super high precision in gears and motors with sensors and image recognition will require a lot of sensors and some pretty hefty processors for image recognition.
High precision hardware means that you can mathematically prove the correctness of motion within a very narrow margin of error.
I think there what you are really trading though is precision in the mechanical components for precision - and speed - in the sensors. If I can detect the slop I can correct for it. If I can't (or don't) then success is dictated by the input, not the output. If you can't detect the slop fast enough you have to slow down to avoid slamming into something fragile or immovable.
The human brain has a motor cortex, uses proprioception as the primary feedback, but touch and sight are used as sanity checks. In my experience with physical talents, you're training your proprioception as much as anything. Probably because it's faster than touch (and has fewer consequences), and cheaper than sight (you can't focus on anything else if you are watching your hands).
The line is getting more blurry as time goes on anyway. There's some extremely flexible hardware out there these days, where most the ongoing development is writing software/firmware for the hardware. That software then programs the HW to do what you want, and so you get the speed benefits of HW with a lot of the flexibility of SW.
Some of the speed benefits. Programmable HW can't match an ASIC in speed yet, but it can significantly outperform a general purpose microprocessor at the same task, provided the task is sufficiently narrow.
Well, if all you have is machine learning, you're going to need a lot of trials. Hence the need for repeatable simulations.
From the article:
"Alternatively, one could follow the Tesla Autopilot approach and deploy their research code in “shadow mode” across a fleet of robots in the real world, where the model only makes predictions but does not make control decisions."
Does Tesla really do that? Has anyone decoded what they're uploading? How much upload bandwidth does each car use? Or is this just hype?
My understanding is that each car can run a shadow model, and then they have a data collection thing running that uploads interesting video clips, where one of the possible conditions is "something different than what we predicted happened". Those clips may then get analyzed, labeled and used for training.
Some Tesla owners seem to think that the cars are learning from all the weird situations they encounter, and I think that's just plain wrong. But some of them, yes.
This reminds me a lot of digital twin and digital ghost[1]. Once you get strong models of those systems then the sky is the limit in terms of training. I could see there being a trend in the future to model and design systems to easily capture and create models of different hardware for companies to help with training development in simulations.
[1]: https://en.wikipedia.org/wiki/Digital_twin
This was a fascinating read (for someone not in robotics).
Beyond the major points, one snippet stood out to me:
> In becoming robust under varied conditions, the simulated policy can treat the real world as just another instance under the training distribution.
Made me think of the Singularity, how we are approaching it from both sides. Humanity, slowly losing grip on what is real, choosing fantasies, manufacturing belief systems - and machines, learning infinite realities, trying to find the truth from data. It's wild.
I think this is a much more interesting angle to the metaverse than virtual reality and 3D glasses. Simulation can also help with personalised manufacturing and designing sustainable cyclical industry.
> Ephemeralization, a term coined by R. Buckminster Fuller in 1938, is the ability of technological advancement to do "more and more with less and less until eventually you can do everything with nothing,"
> Consistent with this theme, I believe the solution to scaling up generalist robotics is to push as much of the iteration loop into software as possible, so that the researcher is freed from the sheer slowness of having to iterate in the real world.
...
> The most obvious way to ephemeralize robot learning in software is to make simulations that resemble reality as closely as possible.