Here is the full run of MicroFactory, that autonomously assembles a photo frame - a real product sold on Amazon. This is a good example of how to deal with complex tasks, while we do not yet have a large set of robotic data to train a big general... ai model. Thread:show more

The key is to break the task down into pieces and fine-tune ai model only for parts where we have randomness. We broke down the photo frame assembly into 27 separate actions: 23 recorded movements and 4 tuned models. For example, when we needed to take a wooden frame from a random stack, we fine-tuned an AI model for this:

After the frame is in the holder, there is no randomness, and the frame can be flipped with just a recorded movement.

Or when you need to carefully peel off the protective film from the adhesive layer by making several scratches, we use tuned model.

But when a tab has been peeled off, the recorded movement is sufficient to finish the task.

As we can see, AI models and recorded trajectories complement each other nicely in real production. Would it be better to set up entire task with robotics AGI? Yes. Could this approach be a step for data collection and be economically viable, enabling robots to do work autonomously using current state of ai? Also yes.

We are preparing a big announcement for next week. Stay tuned!

I have a feeling these are not the best machines to do this task.

Looks like our future is waiting around for slow-assed robots to do simple tasks we don't want to do anymore.

Peeling tape/film without a lift tab is a really difficult problem for a robot

Really cool way of mitigating the speed issues with robot learning models by only using it for tasks that have random inputs. But how much fine tuning is needed per task(like your demo for taking a wooden frame from a pile) ?

If you need to break every step in the same way as programming any other anthropomorphs via a digital twin repeater, what are the advantages of this over any Kuka?

real

Hi Igor. I just read this paper on iterative self improvement for VLAs on very fine grained manipulation tasks. Seems like it would be super relevant to your work:

This is so cool 😇 congrats on the progress so far

This is awesome. It's like a doll house for engineers. I need this

Chez Dassault Systèmes, j'ai vu les projets AI repoussés deux fois en 2023 à cause de la bureaucratie. Impossible d'intégrer ces usines autonomes à bon rythme, dommage pour le secteur en France.

Awesome work! Congrats!

This is awesome!

The design looks interesting

Igor, that's a brilliant example of smart automation! It shows how we can still achieve complex things without needing a massive AI data set, right?

to couple two servos together, were there any special considerations?

That sound of motors is like they are talking.... Naaa naaaaa, na naaaa. Nananaa. 😂

Imagine how fast a single chinese sweat shop worker could do this. Imagine the profit.

cool！

Dope

This is really cool. The speed is commendable

Damn! That's hot! The film cover removing and take off is very smooth.

I really like this idea of decentralised micro factory, cz! How much do you think you will need to make it useful for real customers?

why does it sound like Twin Peaks The Return in the background?

did you guys ever seen a production factory?

That’s some next-level assembly magic! 🛠️

The performance is incredible. It's really impressive to watch. I'm truly fascinated. How did you manage to make it work autonomously? I would have expected two guys with joysticks.

Why are they shaking so much? Is it stiffness or tuning?

Let's say this is 10x slower than the current human solution. This will require 10 of these cells, plus the space for them. @grok estimate the cost of this over time Vs a standard human assembly.

@grok find this on Amazon if it exists there like he says it does

roi: 1000 years

Are you sharing info on the sort of models you’re using and fine-tuning? I often daydream about training a model for some of the stuff I do with robotics but then I always get there without it.

The photo frame is sold on Amazon? I don’t see anything other than a book for micro factory. This is fascinating.

@BrianRoemmele