Comment by momofuku | Hacker News Reader

Comment by momofuku | Hacker News Reader

Having been a member of the robot learning community both in grad school and now in industry, I'd actually like to rightfully attribute something here since it seems that TRI is (deservedly so, I will agree wholeheartely) receiving most of the praise:

The core of these advancements are powered by Diffusion Policy [1], which Prof. Shuran Song's lab at Columbia (before she moved recently to Stanford) developed and pioneered. I'd suggest everyone to view the original project website [2], it has a ton of amazing real world challenging experiments.

It was a community favorite for the Best Paper Award at the R:SS conference [3], this year. I remember our lab (and all other learning labs in our robotics department), absolutely dissecting this paper. I know of people who've entirely pivoted away from their projects involving behavior cloning/imitation learning, to this approach, which deals with multi-modal action spaces much more naturally than the aforementioned approaches.

Prof. Song is an absolute rockstar in robotics right now, with several wonderful approaches that scale elegantly to the real world, including IRP [4] (which won Best Paper at R:SS 2022), FlingBot [5], Scaling Up Distilling Down [6] and much more. I recommend checking out her lab website too.

[1] - https://arxiv.org/abs/2303.04137

[2] - https://diffusion-policy.cs.columbia.edu/

[3] - https://roboticsconference.org/program/awards/

[4] - https://irp.cs.columbia.edu/

[5] - https://flingbot.cs.columbia.edu/

[6] - https://www.cs.columbia.edu/~huy/scalingup/

To be fair, they do credit Professor Song and the paper you linked. TRI is also listed as a collaborator on the paper.

> Diffusion Policy: TRI and our collaborators in Professor Song’s group at Columbia University developed a new, powerful generative-AI approach to behavior learning. This approach, called Diffusion Policy, enables easy and rapid behavior teaching from demonstration.

Interesting that we have a genius robotics Dr. Song for real vs Star Treks Dr. Soong :)

I'm glad that she skipped Dr. Soong's ambition that preceded his work on robotics!

My dad, who worked on military IFF before he retired, met someone in the UK intelligence community whose actual name was James Bond.

Or so he said…

There's actually 1000+ people named James Bond so it's likely. Although fathers are the biggest liars - mine told me he swam across the ocean.

It should be noted that Diffusion Policy (not to mention IRP) was also apparently joint work with TRI.

Can anyone ELI5 (well, or, "explain like I'm someone who understands how autoencoders, transformers & convolutional networks work") diffusion?

What makes it work so much better than alternatives mentioned above?

I haven't read the paper on Policy Diffusion yet so I don't know what they do differently. But I can ELI5 image diffusion models, like stable diffusion. Essentially you add random noise to an image, and the ask the model to predict the noise, such that if you remove that noise detected by the model, you obtain the original image. After the model has been trained enough, int the noise removal task, you can pass just random noise, ask the model to remove noise from the noise only image, then remove a little bit of the noise the model suggested, and do it again. And again, for multiple steps, eventually all the noise is removed and you end up with an image "dreamed" by the model from random noise. You can also condition the noise removal with things like text or other images to guide the noise removal process toward a certain target image.

It seems some researchers in her lab were also involved with Toyota.

> our lab

Which lab are you referring to?

I meant the academic lab that I was a part of while in grad school (would like to keep that anonymous for now, it's a smallish community).

Btw, I was at R:SS 2022 and meeting the Skydio autonomy team was one of the highlights of my career as a robotics engineer!

You may as well credit the information theorists, mathematicians, and physicists who laid out the fundamentals that brought us here.

They died before hardware achieved their decades old visions. Not much of this work is net new description, moreso normalizing old descriptions with observation now that we can actually build the old ideas.

Prompt: ChatGpt write a generic anti academia rant against robotics research

Reply on news.ycombinator.com