Which representations are meaningful for control?
We're presenting TD-JEPA as an oral at ICLR🇧🇷: a zero-shot reinforcement learning algorithm using self-prediction (JEPA) to learn representations that are predictive of long-term, policy-dependent behavior. It works pretty well!🧵
My first PhD paper is out! 🎓
"What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?"
tl:dr: JEPA-WMs for robotics: learn dynamics on top of visual encoders, optimize actions towards goal 👇
w/ @JimmyTYYang1, Jean Ponce, @AdrienBardes, @ylecun
Meet BFM-Zero: A Promptable Humanoid Behavioral Foundation Model w/ Unsupervised RL👉 https://t.co/3VdyRWgOqb
🧩ONE latent space for ALL tasks
⚡Zero-shot goal reaching, tracking, and reward optimization (any reward at test time), from ONE policy
🤖Natural recovery & transition
Exciting PhD position open at FAIR in Paris. We are looking for a candidate to join our team and contribute to advancing the field of AI, especially reinforcement learning.
Find more details and apply below. Feel free to reach out to me by email.
https://t.co/ke2LHai6QI
DINOv2 meets text at #CVPR 2025! Why choose between high-quality DINO features and CLIP-style vision-language alignment? Pick both with dino.txt 🦖📖
We align frozen DINOv2 features with text captions, obtaining both image-level and patch-level alignment at a minimal cost. [1/N]
Honored that our paper Temporal Difference Flows received the Best Paper Award at the #ICLR2025 World Models Workshop, and has also been accepted as a spotlight for #ICML2025! All made possible with the exceptional team @AIatMeta!
📄https://t.co/bMxoJDiH5X
https://t.co/qsn7mYkoR7
@ChenTessler With over-actuated robots, we have observed unstable training, e.g., integration diverging. In such cases, it was too easy for the discriminator to separate mocap and agent trajectories, leading to almost no signal for the learning agent.
Excited to share our latest work at Meta FAIR: introducing Meta Motivo (https://t.co/BBRIFr73m9), a behavioral foundation model for controlling virtual physics-based humanoid agents.
@ChenTessler This approach is good when the objective is to perform tracking, behavior distillation, and then utilize the learned policies as frozen behavioral priors for hierarchical RL.
@ChenTessler There is also a difference in the modeling of humanoids. A few works (like PHC) use humanoids with unconstrained joints and over-actuated controllers.
We're thrilled to open-source the pre-trained model, APIs, and training code, so you can easily integrate it into your own projects. Plus, we've set up an interactive demo where you can see Meta Motivo in action.
Code: https://t.co/6Hksv3EKRR
The result? A model that can solve a wide range of unseen tasks — including reward optimization, goal reaching, and motion tracking — without any extra learning or task-specific fine-tuning, while being robust to unobserved environment perturbations.
How to get in touch: i) send me an email (I would like to chat with you at ICML!) ii) apply online through our website: https://t.co/WzaZA2BWFv.
Please forward this information to anyone in your circle who might be interested in joining our team.
I'm going to ICML next week!
We are looking for a Postdoc to join our reinforcement learning team at Meta FAIR. Our research spans different aspects of RL, with a particular focus on unsupervised/self-supervised learning to build foundation models for AI agents
I will be attending ICML 2023. If you are on the job market and interested in deep reinforcement learning and foundational models, please reach out to me. I would be happy to chat about opportunities related to our project.
🎉Exciting news! Applications for the Reinforcement Learning Summer School Barcelona 2023 are now open! Don't miss your chance to learn from top experts in the field! Applications close on March 27!
Application form: https://t.co/sd33SyPqzF
#RLSS
#ICLR2023 accepted paper:
"Contextual Bandits with concave rewards, and application to fair ranking", by Virginie Do (@gini_do), E.D., M. Pirotta (@teopir), A. Lazaric (@alelazaric), & N. Usunier. Special kudos to Virginie Do. Checkout the preprint here https://t.co/dP7wFWKzGd