Tesla Optimus Sub-Prime Robot

[Buckminster]

Results came in for this poll regarding unboxed:
Snap poll: What is the machine that will build gen 3 that is hard to copy?

Folk seem to think Optimus won't be ready for prime time in two years. Mongo's last post in the unboxed thread I think describes why a sub-optimal Optimus (see what I did there) should still work. Minimal Capex, good use of floor space when you consider it can occur on four plus floors.

So what will Optimus be doing in two years if Tesla don't have the conviction to use him for his intended purpose?

 

[Knightshade]

I think you're confusing "unboxed depends on/requires optimus" with "unboxed will use robots in place of humans where it makes sense at the time"

The first one is what your poll seemed to be trying to get people to say (and resoundingly they disagreed with you)- and is based on a terribly faulty assumption that Tesla would base their next-gen car on hardware they've no idea if/when it'll even work properly.

The second is the likely case, which would potentially offer significant cost reduction if/when optimus can do useful human-replacement- but is never dependent on optimus doing that and they can go ahead and still crank out millions of unboxed cars even if Optimus never proves useful at all.

 

[Buckminster]

https://twitter.com/x/status/1752532999021965444

 

[Buckminster]

https://twitter.com/x/status/1752551650873463092

 

[Buckminster]

Plus Douma is back:

 

[Buckminster]

That's very interesting, I've honestly never considered the possibility that humanoid robots could work FASTER than humans. More hours per day for certain, but faster?

And truly right now I'm wondering why I never considered that before, seems reasonable upon contemplation...

I'm doubling down that there won't be many Kuka's going forwards.:

Check out this old Berlin fly through video:

1. Stamping
   1. Single Kuka required per stamping - no change
   2. Optimus to provide logistics instead of humans
2. Gigacastings
   1. Single Kuka required per casting - no change
   2. No change to logistics
3. Welding
   1. Massive reduction (maybe 90% less to increased size of gigacastings and reduced stampings
4. Painting
   1. No change
5. Assembly
   1. Humans replaced by Optimus

I think each line could go from ~100 Kuka/FANUC style robots per line to ~10.

I can envision an Optimus swinging its torso around from side to side so fast as it works on its tasks ... that it uses so much energy stopping its own momentum... that Tesla decides to add regenerative rotational braking to recover most of that energy. Too out there, or right on target?

Optimus will improve one second at a time. Theoretically they could job share when one robot becomes critical path. I don’t expect there to be a singular line for each module.

The front module could have 3 work centres with ten identical workstations in parallel. If one falls behind, a bot can seamlessly move to another work centre.

It sounds like moving back to pre Ford manufacturing but the flexibility will produce huge benefits. Minimal tooling, seamless human to bot interchangeability and minimal software writing. Imagine how much work improving a Kuka by a second is. Not worth it unless that Kuka is critical path.

 

[heltok]

Another chinese player has entered the game:

https://twitter.com/x/status/1753485104796897473

 

[heltok]

From the comments, I realized that Dreame(the vacuum cleaner winner of 2023 I linked above) also will do humanoid robots:

 

[Buckminster]

https://twitter.com/x/status/1754557288369549380

Marques was able to catch balls and play ping pong

 

[Buckminster]

 

[heltok]

https://twitter.com/x/status/1754564972913332703

 

[heltok]

https://twitter.com/x/status/1755380475118719407

Aloha making some decent gains. I think their robot school is interesting giving me a feeling for how Tesla might set up a school for teaching robots new tasks.

 

[heltok]

https://twitter.com/x/status/1755391360084165094

 

[heltok]

https://twitter.com/x/status/1755645373475869008

 

[Tiger]

https://twitter.com/x/status/1755645373475869008

A whole bunch of nothing there that a few lines of C++ couldn't do.

 

[UkNorthampton]

https://twitter.com/x/status/1755645373475869008

I wasn't sure that families would have a need for an android. Then I saw the block picking up and the picking up of the soft toy.

I'm now convinced that a lego safety bot would be welcomed by many parents wanting to avoid painful late night/early morning lego stepping-ons.

 

[DanCar]

Research shows how using LLMs improves training speed of robots. Nice video on the web page with explanation.

Learning to Learn Faster from Human Feedbackwith Language Model Predictive Control

Project page for Learning to Learn Faster from Human Feedbackwith Language Model Predictive Control

robot-teaching.github.io

 

[Cosmacelf]

Research shows how using LLMs improves training speed of robots. Nice video on the web page with explanation.

Learning to Learn Faster from Human Feedbackwith Language Model Predictive Control

Project page for Learning to Learn Faster from Human Feedbackwith Language Model Predictive Control

robot-teaching.github.io

But the system they use has hard coded programs for each behavior. You’re just using an LLM to write code, which is fine, but hard coding robot tasks isn’t very useful. You end up with a very brittle robot behavior. Unless I’m missing something since I just skimmed it.

 

[DanCar]

>..., but hard coding robot tasks isn’t very useful. ...

How so? If you want a robot to give you a high five, what are you going to do? Tell it not to raise its hand? Yes it can be much more complicated than that, but it is a start. Research is often about breaking down complicated problems into simpler ones, then extending from there.

 

[heltok]

How so? If you want a robot to give you a high five, what are you going to do? Tell it not to raise its hand? Yes it can be much more complicated than that, but it is a start. Research is often about breaking down complicated problems into simpler ones, then extending from there.

Either you tell the robot "function high5() set pose = 0.1 0.4 0.9 1;" or you input that string into a neural network something like this:

and out comes "pose 0.11 0.41 0.88 1.01" which is basically the same, but the neural figured thinks it better and can adapt to the environment.

Then you extend this to include language in the input aka "go fetch me a beer and then take the dog out for a walk".