As PhD application deadlines approach, we are super excited to announce cs-sop.org, created by Zhaofeng Wu Alexis Ross @ZejiangS💻 cs-sop.org is a platform with statements of purpose generously shared by previous applicants to CS PhD programs 🧵(1/n)

Our co-founder and Chief Scientist Drew Bagnell shares the next part of his AI blog series focused on AI alignment and our approach to ensuring the safety of the Aurora Driver. Read the blog here: bit.ly/3WBFGIT

Danke Schön, Vienna! #ICML2024

bumpin’ that

It is officially less than a week before the workshop begins⌛️ The workshop schedule is posted here: rlbrew-workshop.github.io/schedule.html A complete list of accepted papers can be found here: rlbrew-workshop.github.io/papers.html

Come enjoy some brews after RLBReW :)!

I'll be at #RLC2024, helping organize the RL Beyond Rewards Workshop workshop, cheering proudly for our 2 orals at the RL Safety Workshop (rlsafetyworkshop.github.io) and perhaps posting a meme or two on RL_Conference! As usual, DM if you'd like to talk imitation, RLHF, or what's next :).

10th floor!

We had a thriving morning poster session, followed by another one at 3:30. Swing on by to the Marriott Center on the 11th Floor then to participate!

If you enjoyed / missed Jingwu Tang's talk on multi-agent IL (arxiv.org/abs/2406.04219) or Nico's talk on efficient IRL without compounding errors (rlbrew-workshop.github.io/papers/15_effi…) at #RLC2024, stop by the RL Safety / RL Beyond Rewards Workshop workshop poster sessions this afternoon to hear more!

Maybe this alone made RLC worth it 🥹

If you weren’t able to make it to our workshop, the talks are now public!

Cool variant of SPO that learns a latent-conditioned policy for *controllable* generation. Leverages an under-appreciated benefit of preference models: they always produce outputs in [0, 1], making outputs more comparable / tradeoffs more reasonable than across reward models.

Sriyash Poddar Yanming Wan Given latent conditional reward, optimizing policies with this is hard, due to scale ambiguity in RLHF methods. We show that methods like self-play optimization (SPO from Gokul Swamy) can help, since rewards correspond to likelihoods instead of arbitrarily scaled utilities (3/7)

It was a very humbling and optimistic experience to spend a week coding with these high school students. I barely knew how to code at their age, and some of these students were coding up complex search algorithms on real robots in a matter of hours. Thank you CATALYST students!!

The more time I spend in my PhD, the more I find myself agreeing with these points.

ideal PhD application SOP:

Can a single neural network policy generalize over poses, objects, obstacles, backgrounds, scene arrangements, in-hand objects, and start/goal states? Introducing Neural MP: A generalist policy for solving motion planning tasks in the real world 🤖 1/N

Not all prediction errors are made equal! In our new #corl2024 paper, we use the mathematical notion of regret to automatically identify when prediction failures actually led to downstream performance degradation. Website: cmu-intentlab.github.io/not-all-errors/  (1/n)