Figure this is amazing

[Legend]

Tomorrow, Figure will reveal the most significant AI update in our company's history

— Brett Adcock (@adcock_brett) February 20, 2025

He's been hyping it up. Will Figure deliver?

[Legend]

Meet Helix, our in-house AI that reasons like a human

Robotics won't get to the home without a step change in capabilities

Our robots can now handle virtually any household item: pic.twitter.com/Wsx5s8Qelc

— Figure (@Figure_robot) February 20, 2025

https://www.figure.ai/news/helix

This is really really cool.


First off, they have switched to just using a single neural net like Tesla. And they are controlling that net like I've attempted to do with my own robot.

But they seem to have gone much further. The controls of the neural net are not just a few basic things, but they are a rich and complex latent space that is generated by a second model. This visual language model is essentially the conscious brain while the first model is the body mover.


The vlm is 7 billion parameters and the body model is 80 million. The first runs slow, only a few frames per second, while the body one runs 200 frames per second.



I do not know how this compares to Tesla's Optimus, but it is a great development for Figure. Tesla has been a single set of weights since the start since they are building their brain in the same way FSD works. But I do not know how it integrates with language commands or the much richer possibility space of a robot vs a car. I know Tesla is building a super smart chip that will power the next gen of optimus so maybe this is a reason for that. Tesla is probably ahead of Figure here, at the very least their model supports leg movements while Figure's does not, but one super impressive thing is that Figure only trained this on 500 hours of demonstration video. That is practically nothing compared to how much training data optimus needs.

The demo video itself was kinda lame but read the link if you have the time. It's much more interesting. I was worried the demo was all smoke and mirrors since the robots do a dumb look at each other after passing over objects, but it's legit.


edt: they still do the really dumb return to rest pose. I have no idea why that is still a thing, it's good evidence that this approach still is piecemeal in some way and not running from start to finish. I wish they released the full unedited video since you can tell parts have been cut out.