[keeda]

Physical labor, especially jobs requiring dexterity, will be left for a long time yet. Largely because robotics hardware production cannot scale to meet the demand anytime soon. Like, for many decades.

I actually asked Gemini Deep Research to generate a report about the feasibility of automation replacing all physical labor. The main blockers are primarily critical supply chain constraints (specifically Rare Earth Elements; now you know why those have been in the news recently) and CapEx in the quadrillions.

[sumedh]

> Like, for many decades.

Didnt people say that AI is 50 years away in 2010s?

[keeda]

Yeah and until ChatGPT I thought even 50 years was optimistic, which is why current days feel like SciFi! However, at its essence, the current AI revolution has been driven primarily by a few key algorithmic breakthroughs (cf the Bitter Lesson), which are relatively easy to scale up through compute.

On the other hand, the constraints on robotics are largely supply chain-related. The current SOTA for dexterity in robots requires motors, which require powerful magnets, which require Rare Earth Elements, which are critically supply-constrained.

To be precise, the elements are actually abundant in the Earth's crust, just that extracting them is very expensive and extremely toxic to the environment, and so far only China has been willing to sacrifice its environment (and certain citizens' health), which is why it has cornered the market. Scaling that up to the required demand is a humongous logistical, political and regulatory hurdle (which, BTW, is why I suspect the current US adminstration is busy gutting environmental regulations.)

Now there may be a research prototype somewhere in some lab that is the "Attention Is All You Need" equivalent of actuators, but I'm personally not aware of anything with that kinda potential.

[direwolf20]

Some types of motors don't require permanent magnets. If we need more motors than we can make permanent magnets, we'll adapt, perhaps with an efficiency loss.

[keeda]

Motors with permanent magnets are preferred because they are much more cost- and energy-efficient, even with the painful reliance on REEs. There is a very strong incentive to find alternatives but nothing comparable has been found yet.

There are of course non-electric alternatives like hyrdaulic and pneumatic actuators but they are mostly good for power, not dexterity. The size and complicated fluid dynamics simply are not conducive for fine motor control. I do think these will play a large part eventually because even electric motors cannot economically produce enough force to be practically useful. Like, last I checked, the base-level Unitree robots can lift 2kg or so? Not even enough to lift a load of laundry.

At this point I suspect we'll end up with hydraulics for strength (arms, legs, torso) and electrics for dexterity (grippers)

[imtringued]

Uh, out of all the things that are the bottleneck, you think it's robotics hardware that is the bottleneck?

In an age where seemingly every single robot company has a humanoid prototype whose legs are actively supported through high powered actuators that are strong enough to kick your ribs in?

In an age where the recent advancements in machine learning have given bipedal walking a solution that is 80% of the way to perfection with the last 20% remaining the hardest to solve?

Honestly, from a kinematics/hardware perspective the robots are already good enough. Heck, even the robot hands are pretty good these days. Go back 10 years ago and the average humanoid robot hand was pretty bad. They might still not be perfect today, but they are a non-issue in terms of constructing them.

The only real bottleneck on the hardware side is that robot skin is still in its infancy. There needs to be some sort of textile with electronics weaved into it that gives robots the ability to sense touch and pressure.

What has remained hard is the software side of things and it is stuck in the mud of lack of data. Everyone is recording their own dataset that is unique to their specific robot.

[keeda]

Note I didn't say the bottleneck is the hardware itself, it's the supply chain for production of the hardware. Specifically the Rare Earth Elements, as I explained here: https://news.ycombinator.com/item?id=47178210

A bit more detail in this article: https://www.adamasintel.com/humanoid-robots-and-the-future-o...