Tsla Optimus gen 2

Oh, no - we’re very far away indeed. As noted above, we’ve had robots that can pick up an egg and put it down for several years now. Because they’re programmed to do it. A human has “told” the robotic hand what an egg is, how many newtons of force to apply so that the egg won’t fall out or be squashed, and what movements to make. Physically, having robotic hands that can manipulate fragile items has been solved for a while.

Like I said, it’s the brains that are the bottleneck for a general purpose robot capable of emptying a dishwasher. There’s no massive set of data to train an AI on for unloading dishes. There might be some large amount of video of people unloading dishes (though I suspect it’s not all that much), but no data that’s useful for the machine to “learn” the friction and durability and weight of all those objects. Manipulating an object in an environment requires vastly more data than visual records can provide. We’re insanely good at it because we’ve had untold millions of years of evolution refining our nervous and muscular systems. We’re nowhere close to being able to replicate that in a robot.

1 Like

@albaby1 You seem to think that the Optimus will be ready when it will be able to do EVERYTHING. This requires AGI.

I think Optimus v1 production does NOT require AGI. It will be used by Tesla internally when it does just ONE (or few specific) things right that CANNOT be done by today’s robots and Tesla requires humans.

Training a Robot to assemble wiring in a car (or install Solar panels) is a much easier problem to solve than FSD. Today it does require a human and is time and cost intensive.
I think Optimus could be trained to do these jobs flawlessly 24 * 7. They already have the inference-chip/training and Dojo network. They need to get the mechanical stuff right. Other companies are working on POCs in expensive labs mostly while Tesla is focusing on costs and real mass production

Here is an example of the wiring patent they filed

Not at all. I simply don’t think Optimus will be anything worthwhile until it can do something that can’t be done with a conventional robot form factor. Things where the humanoid form factor is useful, rather than an aesthetic.

Auto factories are filled with robots already. They have been for decades. Almost any job that can be performed by robots is already being performed by robots. They’re just not shaped like humans. The reason you don’t have robots installing wiring harnesses is not because we don’t have robots with a human form factor. It’s because it is really hard for robots and computers to manipulate floppy things. If you can solve that really hard problem, you can have a robot install wiring harnesses - and it won’t need to have a humanoid form factor to do it. In fact, a humanoid form factor will probably make it less efficient than giving it the stability of a permanent mount and designing the appendages and body specifically to the task of installing wiring harnesses.

The innovation of a robot like Optimus is in its ostensible flexibility or in its ability to be deployed in environments that can’t be modified to accommodate a robot. It’s useful for being a general purpose tool, rather than a single-purpose tool (like an auto factory robot today), or for being able to operate in human environments (where it might need to climb a staircase or open a non-mechanized door) rather than industrial ones that can easily be modified to meet the needs of robots.

Optimus won’t have “arrived” until it can do something that can’t be done with the type of robots we’ve been using for decades. And working a single job on an auto production line isn’t that thing.

2 Likes

Some perspective on this discussion. Think of looking at the vacuum tube based computers of the 1950s. They filled a room, they took kilowatts to run, all they could do (more or less) was count things from data that had been punched onto cards by humans. When people discussed their future back then, they didn’t even think of the iPhone.

On the optimistic side, now we have iPhones and TVs that play almost all media ever made at a pretty affordable price. On the pessimistic side, it took us about 60 years.

An iPhone is like a humanoid robot. How? Before the iphone we had separate devices for talking to people, taking pictures, navigating your car, buying things, sending written messages. And they all cost less than the iPhone. So why do we all have an iphone and no beeper, no camera, no blackberry, and no Garmin GPS? Because the iphone does all of it, and is small and easy to carry around.

The point of a humanoid robot is not that it will ever replace existing automation. If you have spent a lot of money to build a special case robot to put tail light lenses on a model 3, you are not going to rip that out of the factory and put an Optimus in its place screwing on tail light lenses with a phillips head screwdriver.

BUT… SUppose you are paying humans to clean operating rooms and patient rooms where there is a high chance they will get infected. Train an optimus to push a mop, it can, if you have done your job right, do those things and you don’t have to worry about it getting sick. Have optimuses (optimi?) working around radioactive things. Have optimuses working underwater. Have optimuses working outside space stations. All those things where you currently spend a fortune to try to make it save to have a human do something, that is the some of the low hanging fruit for optimus.

So unloading a dishwasher, that is our test case for whether Optimi are commercially interesting? Were computers commercially uninteresting until 2007 when the smartphone was solved? Computers were NEVER uninteresting. From the 1950s on the applications of computers just grew and grew. Were computers uninteresting commercially until they made it inot the home, or into consumer markets? Absolutely not, they were very interesting when IBM was selling them to companies and governments and R&D institutions around the world to do things that humans had sort of done before, but not cheaply and not amazingly well.

It was the 1980s, maybe even the 1990s before computers made it into the home. 30 years or more from the first computers. So Albaby might be right, might be 30 years before a humanoid robot is unloading the dishwasher for you. But so what? In the 30 years leading up to the dishwasher, humanoid robots would have taken over dangerous, boring, relatively low value jobs more and more as their cost came down and their capabilities went up, at least that’s how it worked with computers. You’ll probaby have an Optimus fixing stuff outside the space station long before one washes a dish or changes an adult diaper.

There’s still no guarantee that TESLA will make a ton of money on humanoid robots. There is no “Wright Brothers Aircraft Company” which raked in a lot of the profits from aviation after 1905. The list of computer companies that have come and gone is impressive. Apple was not a phone maker until 2007, Ou sont les Nokias d’antan?

2 Likes

No. It’s simply an example of the type of job that an Optimus can do which would indicate that that Tesla had accomplished something significantly innovative and groundbreaking in robotics. It’s not the only one, of course. A robot that could provide housekeeping services in a hotel, or replace a UPS delivery driver - those would be rather interesting, and might actually need a humanoid form factor.

Why is that low hanging fruit? Yes, an Optimus could theoretically work underwater (to use your example). But the question is why you would need - or want - an Optimus to do that? The human form factor isn’t optimized for moving in water. We swim using the biology we have, but it’s woefully inefficient compared to creatures that have evolved for water movement. Which is why our underwater robots have propellers and rudders, and not legs for kicking. And why our space robots move on tractors and have specialized appendages rather than trying to walk on legs and use human hands. We already know how to make robots that work in dangerous environments. We don’t need them to be bipedal humanoid in shape.

That’s the question. Not whether robots are useful and beneficial and commercially interesting. Of course they are. Rather, it’s what does a robot need to have - or be able to do - in order for a humanoid form factor to bring anything to the table. Because we already can have robots take over “dangerous, boring, relatively low value jobs” today. What is the benefit of having a humanoid robot? The main ones I see are flexibility (you can use the same robot for a multitude of tasks) and adaptability to a human environment (you can use the robot in environments that are entirely shaped around humans). The exact opposite of working in a factory (or underwater or in space).

2 Likes

The humanoid form factor is useful right now because the world we have built is around that form factor. Sometimes the things we find easy are very hard for a robot to do but I can see a robot being able to do things like accounting and everything that entails, lawyer and everything that entails. I suspect with AI that the white collar jobs are going to be slowly slipping away while the blue collar jobs the robots will find a harder time doing.

Andy

It’s only useful if the robot is capable of doing something in an environment that is built around the humanoid form factor and has to stay that way. An auto plant isn’t that type of environment, because it’s relatively simple to design the plant to accommodate the robots. Indeed, all auto plants do that already. And if the robot is just doing one job, in one place, there’s no need for it to have hands or legs, because it doesn’t need to hold a tool or walk around: bolt it to the floor so you don’t have to worry about balance, and build the tool it needs right into the armature (or whatever).

Remember, the use case for something like Optimus isn’t “can a robot do this better than a human?” It’s “can a humanoid robot do this better than a non-humanoid robot?”

1 Like

Well , if that is the case than Optimus already is winning. Because it is portable and a non-humanoid, bolted to the floor, robot isn’t. It also takes up less space in the factory. That is the problem with bolting things to the floor, it takes up a lot of space and can only be used for that specific job.

Andy

How does that mean Optimus is winning?

For example, being portable and not being committed to just one specific job is only useful if there’s any reason why you would need to move the machine or have it do a different job. But if you always need the robot to do that one job, in that one place, 24/7, then there’s no benefit at all to having the machine be portable or capable of doing something else. Similarly, “saving space” is only a benefit if your factory - or that portion of the assembly line - so so particularly space constrained that there’s any real advantage to having a smaller footprint.

Even if those things were to be useful, they come with tradeoffs. Optimus is small (compared to industrial robots), bipedal (meaning it has to balance), and comes with all the limitations of the human form. It can’t reach any further than a human, only has two arms for holding things like a human, and needs a separate tool like a human - and can only use a tool that’s small enough and light enough for it to carry and manipulate. It might take up less space than the robots in the below video - but is that really an advantage? Almost certainly not for that type of job.

1 Like

ababy1, some people seem to be enamored with a humanoid form factor robot, even though it has so many limitations, as you point out. Do we really need a Jetson’s maid robot?

You said: and I quote.

Like I said a portable robot is much more useful than a robot bolted to the floor. Having a robot that takes up less space on the factory floor is also more useful. If you have ever owned a shop you would realize that having a machine taking up space, that can only do one thing is such a waste.

Andy

1 Like

I don’t think so. I think Optimus exists as a Tesla project for two main reasons.

One, Musk seems to have a very Omni magazine vision of the future world. An Optimus washing a Cybertruck in the driveway of a suburban home would be right at home on the cover. Humanoid robots have often been prominent in sci-fi depictions of the future, so I can see it appealing to him.

Two, Tesla needs something to keep and attract AI talent. The official Tesla line is that FSD is pretty close to done. Which makes it very hard to convince smart people pursuing an AI career to come work for a car and battery, especially a car and battery company that’s going to finish its only AI project within a year or two (allegedly). That’s why Optimus is heavily featured in AI Day, which is as much a recruiting event as an IR event.

TBH, I’m not sure Optimus becomes much more than an afterthought in Tesla when (or if) they “solve” FSD. When Optimus was revealed at the first AI day, it made some amount of sense. At the time, robot seemed like the next real use case for AI, the next task to be solved. In a post ChatGPT world, that’s not likely to be true. Almost all of the first exciting uses of cutting-edge AI are now likely to be disembodied, and Musk has launched xAI so he doesn’t miss out. I think Optimus is Musk’s next solar roof or tunneling machine - a project that struck his fancy and met a business need at one point, but just isn’t worth seriously pursuing.

It’s not just Tesla, Amazon and Boston Scientific also have humanoid robots. I am sure there are more if you really looked around.

Andy

Why? Most factories are filled with machines that are too heavy to move, and only do one thing. Sometimes a specialized machine can do the one thing better than a machine that has to be able to do lots of different things. Sometimes a large, stationary machine can do a thing better than one that has to be small enough to move.

A robot that can move around or do different things isn’t necessarily going to do a better job at a task than a robot that is fixed in place and dedicated to a single task. Unless there’s some benefit to it being mobile, there’s no advantage. And unless there’s a situation where you won’t need the robot to be doing the single task, there’s no benefit to being able to re-assign it to another task. And there are many disadvantages to trying to use a robot that’s limited to a humanoid form factor, instead of having a robot that can be designed in any of an infinite different designs to best suit the work being done.

Boston Dynamics sells robots, but not their humanoid one - it’s purely an R&D project. Amazon is considering using some Agility Robotics robots, in no small part because they’re a big investor in the tech - but having seen some videos of them in action, I have the same criticism (and skepticism) about whether they’re anywhere close to being useful. They’re not doing anything that couldn’t be done better if they had wheeled platforms instead of legs.

Only if the use case for the robot is such that there is value in being able to move. For the welding robots in that clip, there is no value in their moving and, in fact, a value in their being securely bolted in place. Frankly, a humanoid robot would have a tough time doing that job at that pace.

1 Like

The aspect being overlooked is versatility. A human can replace an absent assembly line worker and learn his position in about 30 minutes in most cases.

Robots are extremely specialized. Versatility offers much potential for improvement. Current generation robots may only do one thing. Changing that might be a major advance.

Numerical Control robots are "extremely specialized.’ AI Control robots can share their intelligence ‘over the air.’ There are three dimensions where Tesla’s Optimus robots differ from industrial robots

  1. Control (AI vs. Numerical Control)
  2. Form factor (Mobility in human envirinments)
  3. Internet of Things (IoT)

One needs to understand the utility of the marriage of these dimensions to appreciate the value of AI powered humanoid robots.

The Captain

3 Likes

Sure. But versatility in form factor isn’t particularly useful without a very versatile and flexible “brain.” A robot controlled by numbers (to borrow Denny’s formulation) isn’t all that flexible, so the humanoid form factor doesn’t really add anything. In fact, it’s almost certainly a liability in most instances - if the robot can’t do and isn’t intended to do multiple jobs, it’s better for it to be designed for the one job it’s going to do.

The main deficiency for Optimus is that Tesla is absolutely nowhere close to having a flexible AI brain for a robot. The state of the technology is nowhere close to it. People have achieved amazing things with current approaches AI, but entirely in fields where there exists a massive data set to train on. For information uses, such data sets exist and are enormous. For language, virtually every work of writing has been digitized. We’ve posted literally billions of labeled images online, which enable image recognition and creation software. So when we ask an AI to draft a transmittal letter or draw a picture of a cat wearing a spacesuit, the data sets are vast and deep for training.

No such dataset exists for manipulating objects in a physical world. A robot needs to “know” characteristics of objects like mass, density, distribution of weight, durability, friction of surfaces, inertia, and the like. But there’s no massive database of that kind of information. A robot can’t “learn” how to pick up a box of unknown material with an unknown object from watching people pick up boxes - because the imagery of picking up a box doesn’t convey nearly enough information.

With the current state of AI, without the deep dataset, it doesn’t matter how big or fast your computer is or how good your software algorithms are. That’s why Google can’t help you search your email as well as it searches the (much larger) web. And Tesla has no access to the type of dataset needed to teach robots, and no products that it can leverage to help it build one.

Actually, most advanced CAD software has had that kind of info built into it for over a decade. You can simulate car crashes, for example, using all these material properties. Based on this it would not be that hard to use reinforcement learning to train all sorts of object manipulations without using anything physical. Of course some real world object training would be needed to fine tune.
What robots would have an initial difficult time doing is observing, for example, that a metal surface is moist and thus is more slippery than normal.

Mike

3 Likes