[motohagiography]

Makes sense. The end state appears to be that humans should only be supervising ML that generates goal and outcome based behaviors for robots, and the machines will construct tools to solve problems themselves.

The leap from an AI model learning how to replicate a behaviour (e.g. evolving walking to solve problems https://unitylist.com/p/2id/walking-ai ) to reasoning about it in terms of actuators and physical feedback, to assembling a physical model out of a relatively small list of parts seems like a solvable engineering problem when it is broken out into a pipeline.

Those robot parts are basically a version of mechano with actuators that a model would map a behavior to, and the robots in the article would assemble them. When you look at something like Lego or Mechano as an intermediate representation to construct buildings out of, where all objects made from it are essentially a directed graph of those elements, robots designing and building robots seems like less than 20 years away.

e.g. we could functionally specify to an ML model, "produce a digraph of these element parts that has these degrees of freedom, and then load or derive a model that solves for this outcome within the domain of those degrees, where outcome is 'plug cables into a board' "

[neltnerb]

https://www.mujin.co.jp/en/

This is not manual or bespoke and it has sensors. The videos are incredible and they work in real life already.

This one is it moving petri dishes full of liquid without spilling! This is obviously not being pre-programmed to move along some kind of 1980s style fixed paths for welding parts as Alphabet apparently thinks everyone is still doing. The obliviousness of suggesting that using ML models for robotic control is some unique new idea is really off-putting. Mujin has been around since 2012.

https://www.youtube.com/watch?v=3vleHnx7uug&t=136s

The more the merrier, of course, but just dismissing the state of the industry and claiming you've made a huge technology leap (compared to the 80s and 90s instead of something harder)... ugh.

[yunohn]

> as Alphabet apparently thinks everyone is still doing. The obliviousness of suggesting that using ML models for robotic control is some unique new idea is really off-putting.

Intrinsic/Alphabet are not suggesting they are somehow unaware of easily-Google-able state of the art in ML robotics. They literally used to own Boston Dynamics.

From the post, the second demo of their tech (“Two robots use perception, force control, and multi-robot planning to assemble a simple piece of furniture”), is very clearly much more than “moving Petri dishes”.

FAANG has access to the leading factories in Shenzhen, and heavily utilize robot tech in their HW supply chains.

[kortilla]

> FAANG has access to the leading factories in Shenzhen, and heavily utilize robot tech in their HW supply chains.

Do you know what the N stands for in FAANG?

[yunohn]

This is really boring pedantry, that does not further the conversation. Do you have anything to reply to from the rest of my comment?

[kortilla]

It’s not pedantry. Those companies have effectively nothing in common when it comes to HW.

[fragmede]

It's pedantry. FAAG is not an acronym in common usage (and is uncomfortably close to being a slur), so the more easily understood, less correct word was used instead. To point out that one of the companies doesn't produce consumer hardware doesn't invalidate the underlying point, so what is it, if not pedantry?

[cma]

They probably actually still have DVD handling machines.

[taldo]

The few Mujin videos I watched look a lot like PCB assembly pick-n-place machines. A little bit of computer vision, a little sensing here and there, but overall fairly simple pre-programmed moves, on a pretty controlled environment.

[neltnerb]

If you check out the beginning of the video link (I had it fast forwarded towards the end) you can see that it is doing an awful lot more than that, and in 2013.

A pick and place is 2-axis movement with a suction cup. This is controlling a robot arm with a ton of degrees of freedom and developing paths for moving through all those degrees of freedom without hitting anything and using internal models to do so.

I suppose in some very broad sense it looks similar, but the difficulty of x-y + down is way, way lower than what you're seeing in that video.

[cryvate1284]

It is harder than x-y + down, however I don't think this video is impressive really, having slowed down the video it doesn't look special to me and I did work on robotics/machine vision around that time.

[azernik]

It doesn't seem like Alphabet thinks that? Their ad copy explicitly compares "training times".

[zelon88]

I agree. It appears that Google is trying to pat it's own back in a room full of people who have never seen a modern place of production.

[ta988]

Look at how 5+axis CNC machines work now they are pretty impressive as well.

[zelon88]

Humans design products with a variety of design elements to meet different circumstances.

Look at cars for example. Tell an ML model to "make a car that can drive over rocks" and it will give you a rock crawler with the motor in a location where it won't be easy to fix. Tell the ML model to "make a car that is easy to fix" and it will make a car that is probably unreliable. Tell it to make a car that is reliable and easy to fix you will get a car with no motor at all.

I'm not saying it's impossible, because it obviously is possible. I just think your 20 year time-frame is hopelessly optimistic. What good is an ML model that takes 10 weeks to setup that solves a problem that only takes 2 weeks to solve without ML?