Sadly the debugger is missing many nice-to-have features. I wrote a much improved debugger for Jax you can install here: github.com/JesseFarebro... which will probably give you a much better experience 🙂
3) At the World Models workshop, I'll be giving an oral on a new approach to learning a generative model of successor states through flow matching / diffusion.
📍Peridot 201 & 206
📅Mon 28 Apr 5 PM - 5:30 PM
Check out the paper on arXiv: arxiv.org/abs/2503.09817 with a full thread coming soon 🙂.
2) Arnav & I will be around presenting our work on successor feature matching at:
📍Hall 3 + Hall 2B #572
📅Sat 26 Apr 10 AM — 12:30 PM
Check out the website and paper: arnavkj1995.github.io/SFM/
1) Many of the team members will be around the poster for Meta Motivo at:
📍Hall 3 + Hall 2B #555
📅 Thu 24 Apr 10 AM — 12:30 PM
Don't forget to check out the demo for yourself: metamotivo.metademolab.com and the paper now on arXiv: arxiv.org/abs/2504.11054
Excited to be in Singapore 🇸🇬 for #ICLR25! We’ll present 1) Meta Motivo, a first-of-its-kind model enabling zero-shot humanoid control for any reward, goal, or motion; 2) imitation via successor feature matching; 3) flow matching for generative TD learning of future experience.
As an undergraduate student, taking Rich's course at @ualberta.bsky.social was a defining moment in my academic journey. His work and teachings have shaped the paths of countless researchers, including my own. Congrats, Rich & Andy!
And for categorical representations, @brosa.ca openreview.net/forum?id=dVp... is pretty much canon at this point!
This paper by @harwiltz.bsky.social and colleagues is one of the high points of #NeurIPS2024 for me so far:
arxiv.org/abs/2409.00328
Come by West Ballroom A-D #6404 4:30 PM–7:30 PM tonight (Thursday) to talk to @jessefarebro.bsky.social and me about CALE!
West Ballroom A-D #6704
📅12 Dec 11:00 AM — 1:00 PM
bsky.app/profile/harw...
West Ballroom A-D #6404
📅12 Dec 4:30 PM — 7:30 PM
bsky.app/profile/pcas...
Landed in Vancouver for #NeurIPS. We’ll be presenting a couple of papers with @pcastr.bsky.social and @harwiltz.bsky.social 🧵👇 and stay tuned later in the week for an exciting announcement!
How can you 0-shot transfer predictions of long-term performance across reward functions *and* risk-sensitive utilities?
We can do this via Distributional Successor Features. Our recent work introduces the 1st tractable & provably convergent algos for learning DSFs.
#NeurIPS2024 #6704
12 Dec, 11-2
Distributional SFs: enable 0-shot generalization of return *distribution* functions across a finite-dimensional reward function class
"Foundations of Multivariate Distributional Reinforcement Learning"
#NeurIPS2024 #6704
12 Dec 11am-2pm
neurips.cc/virtual/2024...
Wiltzer Farebrother Rowland
The Atari 2600 console and an image of some of the games used in the ALE
📢 In Defense Of Atari 📢
New blog post in which I argue why the ALE is still a valuable resource for RL research!
psc-g.github.io/posts/resear...
👋
This is how our work on classification vs. regression started, it was folk knowledge in many circles but I had overestimated the extent to which the wider community knew this. openreview.net/forum?id=dVp...
Jax RL work as in writing envs in Jax? If so, each has their place, eg, look at Mujoco in envpool vs MJX, there’s a clear tradeoff point as you increase the number of environments.
👋
Picture of Trinity College Dublin's campanile in the front square of the college. Writing overlayed to indicate that the RLDM conference will take place from the 11th to the 14th of June, 2025. Logos for Trinity College and the RLDM conference at the bottom of the image.
Come join us at Trinity College Dublin for the 6th Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM2025)
Abstract deadline January 15: rldm.org/submit
👋 can I be added as well?
