Becoming Expert in Robot Learning

This is a thought experiment that I am going through.

With the rate of how many papers are coming out every day, and to stay on top, how much do I need to read to develop this taste?

1. How fast are robot learning papers coming out every day? From my arxiv search, there were 2661 papers tagged with cs.RO and cs.AI in the past 12 months, so 2661 / 365 = 7.3 papers per day

• I will not keep up with 7.3 papers a day

The good news is that how many of those papers are actually useful though? Not that many.

• For example, this paper self-claims to be groundbreaking… any paper that claims to be groundbreaking is a huge red-flag. That is up to the readers to judge.

First, you can quickly filter out just by looking at school. The top robot learning is really consolidated in 2 schools: stanford and berkely. So you can just read those papers. And then, there are extra benchmark papers like Bridge (lol those also came out of stanford/berkeley).

So action plan:

• RoboPapers is a pretty good signal since they hand-pick the authors
  • ~20 papers right now

I think if I read ~100 papers deeply, I can get a really strong grasp of the field. I think future lifetime goal is 1-2 papers a day. You will be able to read more papers faster once you can quickly identify the new ideas. For example, for you, Flow Matching is a really new idea, so the math doesn’t make sense. Perhaps it’s worthwhile to skip the math.