Reinforcement Learning Kinda Works Now

A collection of links from the Twitter thread:

https://serl-robot.github.io/ "SERL, our open source software framework that enables learning manipulation skills in 15-30 minutes from raw images"

https://sites.google.com/view/reboot-dexterous "dexterous hands doing in-hand manipulation, from pixels, in just a few hours."

https://sites.google.com/berkeley.edu/aprl "just 5 minutes (!) to learn to walk from scratch"

https://sites.google.com/view/fastrlap "Learns in as little as 20 minutes to race indoors, outdoors, etc."

In case this sparked anybody's interest, I went through the tweet's author's course on RL [1] recently and it was extremely insightful. I heartily recommend it (slides, lectures, assignments are all public)

[1] https://rail.eecs.berkeley.edu/deeprlcourse/

Do you know how does it compare the coursera RL course?

Help us out, what was there to notice in the last year?

This Tweet says something I want to be true, but it doesn't help me determine whether it is true and it doesn't help me learn.

In the example shown in the video, what algorithm is being used? What reward function? What exploration strategy?