While responding to a SA comment I realized that deploying Optimus Robots will resemble how IBM deployed computers when I started programming at the IBM Service Bureau in 1960.
One of the things that naysayers are ignoring is that Tesla will put a thousand robots to work at their own factories where they can be debugged without worrying about customer feedback. By the time Optimus is ready to be released to the world it will be quite capable. Tesla will be able to select customers that have jobs apt for Optimus. It’s not as if Optimus can conquer the world in a day. Optimus will need to be trained for every new job. It will be a novel and unique way to release a new product. People who are thinking in terms of industrial robots are missing the point.
In 1962 I was promoted to Systems Analyst and my main job was to install the computers IBM sold. Today you can order a computer a thousand times more powerful and have it up and running hours after getting it. Back then my job was to teach the client’s staff to program and to help them write the code that would run on their new machine. I would take them to the IBM Service Bureau to test and debug their code. I don’t recall how long it took to deliver a computer but it was between three and six month. I recall one case when IBM had to hire a crane to deliver the machines through a second or third floor window.
The point is that Optimus is capable of doing multitudes of jobs but first it must learn to do them! One way would be to control them with tele-operators. After a few days or weeks it would have mastered the job and ALL his brethren would be able to do it. Tesla is learning this skill with the robots it is installing in their own factories.
If this actually happens, then sure. If Tesla is able to develop a robot brain that is capable of mastering a new and useful job with only a few days or weeks of teleoperating, then it will have achieved something enormously valuable.
But that’s the trillion dollar gating issue. Is Tesla (or anyone) anywhere close to developing a robot brain that can do that?
Naysayers aren’t ignoring that Tesla has factories that they can try the robots out in - they just don’t think that’s the relevant bottleneck. You can test robots without doing it in a real factory. It would be trivially cheap to rent a 40K s.f. warehouse somewhere and fill it with industrial equipment and debug the robots there. Building a test factory would be a trivial cost, a rounding error compared to the budget for developing a humanoid robot AI that could master a new job with only a short amount of teleoperating.
The hardware is relatively easy to build. It is the software (datasets plus trained models) that is much much harder. Summary: software: hard, hardware: easy.
There are thousands of robots in all of the factories now. They just do not get up and go to some other corner of the factory. The jobs are all the same regardless.
True - but no workers’ job in a Tesla factory is just to serve you a cup of coffee (or rarely anywhere else). Nor is it apparent that Tesla’s yet developed a “brain” that can master a new job like serving someone a cup of coffee after only a few days or even weeks of teleoperating.
Again - anything to show that they’re there yet?
What’s wrong with the above? It discusses trivia. It entirely misses the new paradigm. Up to now the so called robots were mission specific numerical control machines (NCM) that are very efficient doing the one task they were designed for. Tesla at Fremont found out that it can be overdone so they redesigned the Fremont production process trashing many superfluous robots. The paradigm shift with humanoid robots is that this machine is not purpose built for one specific job, it’s designed to do many tasks humans do in human environments. Effectively humanoid robots learn to do new jobs with an over the air update. It is this paradigm shift that the above conversation entirely misses.
Paradigm shift
noun
a fundamental change in approach or underlying assumptions:
ORIGIN
1960s: term used in the writings of Thomas S. Kuhn (1922–96), philosopher of science.
To understand humanoid robots examine the paradigm shift from Numerical Control to Neural Network AI as the governing entity of the machine.
Tesla is not alone in the game. The first versions of Optimus had rather rudimental hands probably to save on hardware costs. Figure explained why human like hands were of utmost importance. Watch the hand!
Elon Musk recognized his mistake and the new hand has 22 degrees of freedom
Humanoid robots don’t have wheels but could learn to skate.
Sure, because “the above conversation” was constructed by you, rather than reflecting the main skepticism about humanoid robots.
The “naysayers” aren’t skeptical about humanoid robots because of their hands, or because they don’t have factories to train in. Rather, it’s that existing technology is nowhere advanced enough to allow humanoid robot brains to “learn” new jobs in the real world (ie physical manipulation of objects to work in a 3D environment) the way you describe.
It’s not hands - it’s brains. Hands are an insanely flexible tool, allowing a robot to do a wide variety of different tasks, rather than a single one - but that’s only useful if the robot’s brain is advanced enough to do lots of different tasks. It’s the brain, not the hands, that’s the gating technology.
When I noted upthread that if Tesla had developed a robot brain that can master a new and useful job with just a few days or weeks of teleoperating, you said that they already had. I don’t believe that’s true - they don’t have a robot brain that can master a new and useful job with just that amount of teleoperating. Which is why I asked for an example.
The brain is operational in millions of Tesla EVs, FSD. Optimus uses the same inference chip as the EVs. The training is done by neural networks just like LLMs and FSD. The initial data is generated by tele operated Optimus robots. This was just brushed aside saying hardware is easy, software is hard.
Figure disagrees and Elon took them seriously. I remember watching the Figure presentation which made a lot of sense, wondering what Tesla would do. Elon did what good executives do, he corrected his mistake ASAP.
That’s not what’s relevant. The question isn’t whether there exists a type of brain or AI or algorithm. After all, there’s millions of robots today that are run by software.
What they don’t have is a robot brain that’s capable of doing what you describe above: master a new and useful job with only a few days or weeks of teleoperating. There’s no indication that the neural nets they’ve created, or the training data they have, is capable of producing a robot brain that can do that.
The training data for FSD just isn’t useful for any workplace scenarios. Having billions of hours of video of driving won’t be useful for teaching a robot how to pick up a screwdriver and hang a picture (for example). So no matter how amazing FSD might be, it can’t just be ported into Optimus, even if it uses the same chip - because the “learning” isn’t in the chip, it’s in the data and the neural net that was trained on it.
So again - I’m pretty sure that Tesla has not already developed a robot brain that’s capable of mastering a new job with only a short amount of teleoperation.
And wheels would work in some environments … not space, though where one might need little jets. Point being that environments currently occupied by humans can be navigated with legs, so legs makes sense for a robot intended for use in an environment currently occupied by humans.
