←
Raul Steleac21 posts
Raul Steleac avatar

Raul Steleac

@steleac

PhD Student @EdinburghUni; studying temporally extended behaviours in both single and multi-agent RL

◷ Joined August 2019▧ 2 Photos and videos
Raul Steleac avatar
Raul Steleac@steleac·Apr 20

Really excited to present our recent work at #ICLR2026 this week! We discover highly coordinated joint behaviours and integrate them into the skill sets of MARL agents, accelerating the search for effective joint strategies in downstream tasks.🧵 Paper: raulsteleac.github.io/iaro

Post attachment
Leo Hinckeldey avatar
⟳ Raul Steleac retweeted
Leo Hinckeldey@LHinckeldey·Sep 10

A good reason for my first post! A few weeks ago, our paper Assistax received an Outstanding Paper Award at #RLC2026 🎉 Assistax is a GPU-accelerated RL benchmark for assistive robotics. Huge thanks to @RL_Conference for the recognition! More below 👇

0:12
Samuel Garcin avatar
⟳ Raul Steleac retweeted
Samuel Garcin@SamuelGarcin·Jun 16

We’re open-sourcing the models, weights and training code for PERSIST, our 1.7B parameter 3D world model! By modelling a dynamic 3D world state instead of relying on pixel histories, PERSIST generates experiences that remain spatially and temporally coherent over thousands of steps. Tom, @kaixin20578389 and I will present PERSIST at @icmlconf in Seoul next month. See you there! Links in thread⬇️ #WorldModels #ICML2026

0:12
Raul Steleac avatar
Raul Steleac@steleac·Apr 20

Work done under the supervision of Mohan Sridharan and @dabelcs (many thanks)! If you’re at #ICLR2026 in Rio and want to chat, come find me during Poster Session 2! 🔥

Raul Steleac avatar
Raul Steleac@steleac·Apr 20

Finally, we use this multi-dimensional n-distance as a state representation for eigenoption discovery, leading to coordinated alignment patterns that are effective in aiding teams of agents in multiple downstream tasks. Also works with heterogeneous agent state spaces.

Post attachment
Raul Steleac avatar
Raul Steleac@steleac·Apr 20

Having a scalar describe the (mis)alignment of multiple agents is limiting, as it conflates the impact of different state features. We thus propose an information theoretic objective that disentangles the n-distance approximation across each individual state feature.

Raul Steleac avatar
Raul Steleac@steleac·Apr 20

We first generalise pairwise state distances to involve n elements by introducing n-distances in MDPs via our Fermat inter-agent distance. To define this metric, we introduce the Fermat state, a fictitious state of maximal alignment accross agents, and show how to approximate it.

Raul Steleac avatar
Raul Steleac@steleac·Apr 20

To discover such joint behaviours (options) in an unsupervised manner, we follow the intuition that: in the absence of an explicit objective, a natural basis for coordination among a group of agents can be achieved through patterns of alignment of their states. 🤔

yobibyte avatar
⟳ Raul Steleac retweeted
yobibyte@y0b1byte·20 Apr 2022

A lot of people complain that RL doesn't work and RL researchers are still playing games. While this criticism is true to some extent, there's been a new trend of applying RL for real-life problems. This is a thread of notable papers split by the topic. 1/n

Scott Bryan avatar
⟳ Raul Steleac retweeted
Scott Bryan@scottygb·14 Jun 2021

23 and 24 year olds able to book their NHS vaccine appointments from tomorrow.

Google DeepMind avatar
⟳ Raul Steleac retweeted
Google DeepMind@GoogleDeepMind·26 May 2021

Discover how WaveNet has evolved from research concept to advanced real-world system that creates more natural-sounding speech and helps @Google unblock communication barriers for millions of people around the world: dpmd.ai/wavenet

Post attachment
Aran Komatsuzaki avatar
⟳ Raul Steleac retweeted
Aran Komatsuzaki@arankomatsuzaki·12 May 2021

Diffusion Models Beat GANs on Image Synthesis Achieves 3.85 FID on ImageNet 512×512 and matches BigGAN-deep even with as few as 25 forward passes per sample, all while maintaining better coverage of the distribution. arxiv.org/abs/2105.05233

Post attachment
Karol Hausman avatar
⟳ Raul Steleac retweeted
Karol Hausman@hausman_k·19 Apr 2021

In addition to MT-Opt, we are releasing Actionable Models, which addresses the problem of defining tasks (which becomes quite cumbersome at scale). This work uses the dataset collected by MT-Opt but uses goal-conditioned offline Q-learning to learn a general goal-reaching policy.

Excited to present our new work on Actionable Models, an approach for learning functional understanding of the world via goal-conditioned Q-functions in a fully-offline setting! paper: arxiv.org/abs/2104.07749 website: actionable-models.github.io youtube.com/watch?v=S3SCR7iY…