OK, I give up. 
The Captain
2 Likes
I’ll go back to the well.
Capability is not the same thing as execution.
Execution of a simple task requires linear capability for motion, sensing, switching and feedback
When you stack that up, however:
Execution of a simple task type W
Execution of a simple task type R/Q
Execution of a simple task type A
Execution of a simple task type B/C
Execution of a simple task type F/N
Execution of a simple task type A/D
Execution of a simple task type C/D
Execution of a simple task type C/H
Execution of a simple task…
Execution of a simple task…
Execution of a simple task…
Eventually the simple linear capabilities overlap to the point that execution becomes a given REGARDLESS of the next task.
Wire harnesses are an interesting case for how you describe. Did you know that most of the reason we have wire harnesses today is because it’s too complex for a simple human to get it right without it?
The ideal wiring harness is run point to point with the least amount of material and time (and connectors!). How do you think that process will change once we start letting the AGI optimize design and installation paths?
Involving the robot in the process is a key parameter for the very design of the process itself. (and the product!)
2 Likes
one more thing on harnesses.
robotic automation of a harness is as simple as identifying the hard points and the connectors. with simple cantags
(by the way, the connector itself with it’s orientation, coloring, numbering and wire colors is, itself, a marker for machine vision systems which are integral to this process)
The robot can identify where to grab, how to hold and where to place.
This, of course, is just a 1/2 measure until the entire process can be optimized for digital interfaces and ML.
Don’t give up! Maybe you could explain. How do you think a robot could “learn” how to install a wiring harness with a small amount of video of a human doing it? Tesla wouldn’t need Dojo at all for this task, since there are lots and lots of computing options that are capable of handling the small amount of data that would be generated by just one robot creating video.
Of course. The question isn’t whether a robot can do those things, but what you need to do to “teach” them to do them. We have countless robots that are programmed to do tasks that involve identifying where to grab, hold, and place - automakers have been using them for decades.
The question is what kind of processes and data you would need in order to “teach” a general purpose robot to do those things, rather than program them to do it. How many examples of humans manipulating objects with those cantags will a computer need to analyze before the correct patterns emerge from the data? Can they do it just from video data, or do they need data on how much force and tension and pressure is being placed on each component to get the desirable outcome? Etc.
Well, aren’t we getting deep into the weeds, here!
If you are familiar with the rotation and torque procedures for basic fasteners, you should be able to make the jump to displacement and force procedures for installing clips and connectors.
validation then verification component by component is the way. Assembly work instructions are simple to aggregate once each step is understood. That verification part is key because in a multivariate system, the “operator” needs to know to try again.
You see this for high speed automated sequences where rack and rerack are part of the process.
For humans, a great example is how poorly the lug runner does on a tire change when he misses one. His feedback is so poor, and delayed, he has no chance of recovering.
However, a machine with torque and rotation feedback (or displacement and force) is sampling 1000x/second and “tries again” before it even goes to the next decrement.
Machine learning is actually great here because it learns or it doesn’t. At a set threshold, a design or process change to “step” the process up in fidelity is needed. This in turn allows for improved training or, perhaps, an additional micro capability that was not known to be needed before.
When a machine learns like this, it might just reposition it’s sensor or camera instead of throwing a flag, halting and waiting on a fleshbox to come and save it.
I’m speaking theoretically here because, while all of these things are currently possible in semiautomatic STATIC processes, the human still has to close the gaps. We don’t automate to this level today.
Not because it’s not possible, but because it doesn’t make economic sense to do so with so much available human interaction.
On the moon, at the bottom of the sea, inside the reactor core? maybe so.
3 Likes
Sure. Nobody is arguing that with enough experience, trial, teaching, and input a robot can’t be programmed to do a task, almost any task eventually. But there are millions upon millions of tasks that a general purpose robot will have to do to achieve what is being predicted.
“Walk to the car” becomes walk down the hallway, step over the broomstick that wasn’t there yesterday, open the door handle that’s sticking because of the humidity, don’t let the cat out when you open the door, and so on. “Cook a fried egg” which takes you a millisecond to process takes hundreds of steps for the robot - and yes that too can certainly be done.
It’s the “general purpose” that makes it so difficult, and I expect Musk will, at some point, say “Harder than I thought” - as he has with self driving, with Starlink, with Boring, and most obviously with Twitter.
I take it back. There are trillions of things the robot will need to do, and breaking them down into discrete steps (as you indicate) will do a lot of it, but like cars that drive into wet concrete, or don’t see a truck heading at them, or stop in the middle of the road for no apparent reason there are always going to be edge cases and confusions and things that gosh darn it just don’t work.
Someday, just not soon.
Ok Albaby I am asking you to explain how they will do this?
Andy
I don’t think they will. Or at least, not in any relevant time frame, with the technology that we have today.
I’m not sure why Tesla is trying to develop a general purpose humanoid robot.^^ Sure, I get why they would love to have developed a general purpose humanoid robot - you’d make a killing if you had a cheap robot for sale, like in the movies. But the same is true of other “movie” tech. You’d make a killing if you had a working hoverboard or replicator that you could sell at a reasonable price. However, everyone kind of knows that you can’t actually make those at economical prices with current technology. It doesn’t make sense to put a ton of money into trying to bring one to market as a commercial product.
I think the tech just isn’t there to develop a general-purpose humanoid robot.
Albaby
^^I mean, there’s a cynical answer, of course. Tesla had to spend a billion dollars to build a super-computer to “solve self-driving,” while simultaneously saying that’s a problem that will be solved within the year. Actually, when Tesla announced they were starting Dojo in August 2021, Musk was then predicting FSD would be finished faster than they (ultimately) would end up building Dojo. You get a lot of awkward questions on the investor calls if you’re planning to spend a billion dollars on a computer for a problem that’s going to be solved before the computer’s built. So they kind of have to have a “next” use case for that hardware - and to keep the team of AI and machine learning experts they’ve hired staying on, rather than telling them that most of their jobs would be finished before the new toy was built. So…super-cheap general purpose robots, why shouldn’t that be a thing, I ask you?
That is the easy way out. If you can’t imagine how they would do it, it seems you do not have a grasp on the technology. After all you should be able to see both sides not just one side. I can’t see how they will do it but I think, with AI, if it is really as transformative as they say, we could see many things that we didn’t think possible. After all, do you think people crossing the prairies in a covered wagon could see planes flying in their lifetime?
Andy
3 Likes
Just as harnesses emerged to simplify the job for humans Tesla is making rigid harnesses that even robots can manage to install.
The Captain
Not in any practical sense. People crossed the prairies in covered wagons in the 1840s to 1860s. (After that they took the train). The first flight was in 1903, but the Wright Bros submerged for a decade fearing patent copying, and the first common flights weren’t until WWI. So at least 50 years, maybe 60.
We do, after all, have “humanoid robots”. We’ve seen them “walk”, “fall”, “climb stairs”, and so on. What we haven’t seen is any generalized capabilities, nor any economic path to make them viable in general circumstances. There is just too much electro-mechanical complexity, and not nearly enough AI or data storage available or power storage on such a unit to make it work.
I don’t think Albaby or I are saying it can’t ever happen, it’s that we’re nowhere close to that time. Even with the way things go faster now there are myriad economic, technical, and data centric issues that put it in the “hoverboard” category for the near future.
A perfect example, thank you. They’re not solving the “robots can’t handle wire harness issues by improving the robot”, they’re changing the harness . Does that seem a viable solution for everything else a robot can’t manage in a human world?
A perfect example of how pragmatic Tesla culture is. The perfect is the enemy of the good (enough).
Wrong question. Tesla’s current business is EVs, they cannot wait a decade or so to have perfect human robots so they create a pragmatic compromise to keep making EVs and to train robots to do useful tasks.
Optimist is not designed to imitate Boston Dynamics robot antics. At least the first generation will be just a limited version of humans in action.
About data storage, one important difference between human intelligence and AI is precisely about storage. AI relies on huge off-line storage to train neural networks. Once they are trained, a much smaller amount of data needs to be uploaded to the robots. Humans, by contrast, have to do it all in one brain.
Further, no human can do everything. No humanoid robot need to be able to do everything. They will have a minimum common intelligence and sets of specialized intelligence, you know, EV PhD, Nurse PhD, Companion PhD, Clerk PhD. The perfect is the enemy of the good (enough).
The Captain
1 Like
I thought it was clear in my post that I can imagine how they might do it - just that I don’t think they will do it.
In order to use the technology that powers recent publicly prominent advances in AI (and Dojo as well), Tesla would need a massive amount of data to feed in to the learning program. That’s what enables these technologies to “learn” the outputs that we want from them. They can create verbal or picture outputs because we’ve created massive public databases for them to train off of - virtually every piece of English-language writing has been digitized on the internet, as well as billions and billions of pieces of visual art (photos, paintings, drawings, etc.). Dojo can create car driving outputs because Tesla has collected literally billions and billions of miles of both video and vehicle telemetry data to input.
If you wanted to use that type of tech for a general purpose humanoid robot, you’d need to generate an appropriately large dataset to train off of. I don’t think Tesla can, or will, create such a database. Unlike cars, they’re not already manufacturing a product that can collect that data for them. Unlike verbal and visual arts AI, there isn’t a public, free, and massive collection of that data on the internet for them to scrape. Partially that’s because workplace environments don’t get posted to the internet nearly as often as other works, but mostly because video data isn’t enough to “teach” a robot how to perform tasks in the physical world. Just as with cars, Tesla needs additional data about the forces involved in order for a robot to “learn” how to (for example) pick up a cardboard box with contents of unknown weight with enough force to keep the box from slipping out of its hands but not too much force to crumple the box. All the videos in the world of people picking up cardboard boxes would not suffice as training data for that task.
There’s no viable way for Tesla to create a large enough dataset to make a Big Data machine learning process viable to teach a general purpose robot. You could, in theory, have Tesla’s workers wear a ton of haptic and other sensors while being continuously monitored for a sufficiently long period of time - but that would only provide data useful for teaching robots how to do those specific auto manufacturing jobs. Unlike the internet, there’s no incentive for anyone else to voluntarily record that kind of data and make it available to Tesla for free. Unlike Tesla’s cars, Tesla doesn’t make and sell any existing products that are capable of collecting that data for it. And unlike driving (or the internet), most of the observations you need to make are done on private property, out of public view, in workplaces that employers will have a reluctance to share data on.
And more broadly, not all things that we can imagine are possible with our technology. People have been thinking about things like flying cars and underwater cities and moon colonies - and humanoid robot servants - for almost a century now. They still haven’t come to be.
People were still using covered wagons into the 1890’s to get across certain parts of the states. Like everything, it gradually declined. The first plane came out in 1903. I know you don’t think we are close but like every prognosticator that is just a guess. Humans are really terrible at predicting the future.
Andy
2 Likes
I’m not sure what the answer is. This is because history has shown that humans are VERY adaptable (as any species that continues to survive is). And there may be two realms of problems that need to be solved:
- The realm of “created things”. Anything that humans created, whether physical or process or social, can change in ways the robots can deal with it. And usually in relatively short time (single digit human generations, let’s say).
- The realm of humans themselves. Anything inherently human, the robots would have to adapt to. For example, elder care, the robots would have to adapt to the human because the human can only adapt over many many generations (hundreds or thousands or tens of thousands of human generations) if necessary.
So the parts and construction of automobiles (or anything else) can be adapted as necessary to the robots. But the personal care aspects of robots cannot.
1 Like
I think people are trying to make this to complicated. They already have robots performing jobs that are not that complicated, so to say they can’t do it has already been proven wrong. They will just move up the chain as they get better and smarter. How long it will take for them to get to where they are all over the place will be decades but gradually they will get there.
Andy
3 Likes
That’s the pragmatic solution that some take as insurmountable hurdles.
The Captain
1 Like
I don’t think anyone disagrees with that. Robots will get gradually better and better, and over the next several decades they will be increasingly ubiquitous.
That’s a different claim than asserting a general purpose humanoid robot can be economically developed within a short enough time frame to be meaningful to Tesla as a company.
1 Like
This is what I’m arguing. And Albaby too. These dreams are really cool, I’ll admit, but pretty unrealistic - like several others I’ve seen lately. Things will certainly change, just not on these types of gee-whiz time scales.
CompuServe was founded in 1969, the Internet got going in the 90’s. The first cell phone was 1973, the first candy bar phone in the 80’s, flip phone in the 90’s, Jesus phone in 2007. These things play out over time; heck Tesla took five years to produce the economically unviable Roadster and a decade to produce the first successful EV (and that only with a $500B gift from the government.)
Musk may see the future, but it’s very far away. Boring hasn’t done much, a Las Vegas showpiece project that nobody’s copied? He’s at 1/10th the projected revenue and subscribers for Starlink. I am highly skeptical that people will give up their cars in favor of Robotaxis (which doesn’t mean it won’t be a business, it just won’t end private ownership as several here predict.) I am likewise skeptical of the prediction that cars will look like toasters, that we will colonize Mars, that hyperloop tunnels will transport people at 750mph, or that the way to run Twitter is to fire 75% of the staff. (Or, frankly, that there is a path to turn X into a financial & “everything” app - that is as widely successful as he imagines.)
Times scales. That’s what the argument is, not whether or not something might eventually, someday, sooner or later, in some indeterminate time to come, actually happen.
2 Likes