If you are not familiar of the robot-learning/RL papers, and if you have the bandwidth to go through to them, I'm more than happy to share my source papers! Please let me know either ways!

And that some intervention is still needed for robot learning pipelines to demonstrate respectable ICL?

The 8th thread graph makes me very curious, and I'd love to hear your thoughts on this phenomenon!

i.e. converting robot joints into discrete cosine transforms seems to significantly improve sample efficiency and generalizability.

Why do we see this happen? Is that merely a ephemeral local minima that were stuck into?

However when SOTA papers use transformers for learning policies and Q functions in RL, the observation seems to be that "Distributional RL works better for offline RL tasks", or Frequency action space tokenization.

Hi Daniel, great post!
I'm curious what do you think about continuous control use-cases? From my very quick read over the post (not the paper itself) it seems that ICL emerges as a property of the model being able to handle diversity of continuous variable regression.

1/n

I like scholar-inbox
www.scholar-inbox.com

Store it on an old school magnetic HDD!
You only gotta power it on every decade or so to maintain the hardware.

RL as a refinement tool has been used in dexterous manipulation for some time!
It used to be quite hard to do tabula rasa learning for dexterous manipulation. And still is, for the most part!

Hard agree. Although how to reconcile that with when you're writing a paper?

Like what do you think is a reasonable process to pick "baselines"?

I've seen some students get jaded cos reviewers ask them to incl "baselines" with shitty git implementations :(

Hi! Can you please add me to the list?
Thanks for making it!

🙋‍♂️👋