And a few works w/ some of my collaborators:
@DannyDriess
will present PaLM-E, a language model that can understand images and control robots, Tue 2 pm poster
#237
:
Video here:
Students from RAIL (and collaborators!) will be presenting some exciting papers at
#ICML2023
, from state-of-the-art methods for leveraging offline data for fast online RL to new methods for running RL on passive data (e.g., videos). A thread and summary 👇
Tue 5:30 pm,
@its_dibya
presents Intention-Conditioned Value Functions (ICVF), oral in Ballrm C, poster Wed 2 pm (
#202
): learn how we can train value functions w/o rewards on data w/o actions, by learning latent intentions!
Video:
Tue 11 am, see poster
#218
on RL w/ Prior Data (RLPD), a super-efficient RL method. We use this for most of our online RL robot experiments these days! By
@philipjohnball
,
@smithlaura1028
,
@ikostrikov
Check out code here:
Video:
Wed 11 am, see poster
#633
by
@qiyang_li
&
@simon_zhai
on how curricula and conditional policies can provably accelerate RL, w/ polynomial sample complexity w/o exploration bonuses! Paper:
Video:
Thu 10:30 am,
@seohong_park
presents Predictable MDP Abstractions (PMA): learn policies that make predictive models easier to train! This provides for unsupervised RL pretraining, making model-based RL practical.
Video:
Thu 1:30 pm,
@ben_eysenbach
presents work analyzing connections b/w actor and critic regularization in offline RL, to study how the two might be connected and how these observations shed light on both types of methods:
Video:
Also Tue 2 pm (poster
#201
) is a presentation on Jump-Start RL by Ikechukwu Uchendu et al., which describes how rolling in with a pretrained policy can "jump-start" efficient online RL:
Video:
And Wed 4 pm,
@5kovt
&
@ARGleave
will present learning adversarial policies to fool game-playing RL agents at Go, oral in Ballrm A, poster
#300
at 10:30 am on Thu:
Video: