🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

We introduce APT-Gen to procedurally generate tasks of rich variations as curricula for reinforcement learning in hard-exploration problems. Webpage: kuanfang.github.io/apt-gen/ Paper: arxiv.org/abs/2007.00350 w/ Yuke Zhu Silvio Savarese Fei-Fei Li

We are organizing the #RSS2021 Workshop on Visual Learning and Reasoning for Robotics (VLRR)! Submissions are now open! (Deadline: June 20th) Website: rssvlrr.github.io

MuZero removed simulators in MBRL vs AlphaGo. VQ Models for Planning generalize to partial observable & stochastic environments. How? 1. Discretize states w/ VQVAE 2. Train a LM over states 3. Plan w/ MCTS using the LM Led by Yazhe Li & Sherjil Ozair arxiv.org/abs/2106.04615

Nominated for Best Paper at #CVPR2021! Very proud of Jennifer J. Sun! Check out all the nominees here: cvpr2021.thecvf.com/node/290

4K version of my ICLR keynote on #geometricdeeplearning is now on YouTube: youtube.com/watch?v=w6Pw4M… Accompanying paper: arxiv.org/abs/2104.13478 Blog post: towardsdatascience.com/geometric-foun…

Excited to present our work on Actionable Models at #ICML! Find the camera-ready version at arxiv.org/abs/2104.07749 In this work, we learn functional understanding of the world through goal-conditioned Q-learning and use it for reaching visual goals or learning downstream tasks.

Thinking Like Transformers RNNs have direct parallels in finite state machines, but Transformers have no such familiar parallel. This paper aims to change that. They propose a computational model for the Transformer in the form of a programming language. arxiv.org/abs/2106.06981

Using a diffusion kernel to represent the neighborhood structure in a graph for processing by a transformer. Cool stuff.

A kind of GAN for system ID: Train simulator parameters so that the simulated *trajectories* are indistinguishable from real trajectories, treating the simulator as a "policy" that tries to replicate long-horizon behavior seen in the real data. arxiv.org/abs/2101.06005

Brilliant! Differentiation strikes again. Question/suggestion: same trick for the selection in evolutionary algorithms like PBT? ie differentiable natural selection?!

Sharing Yoshua's important post (from FB) for the AI and science folks here... jacobbuckman.com/2021-05-29-ple… Jacob Buckman

Very pleased to share a new pre-print with Jose Miguel Hernández-Lobato and Yarin OATML_Oxford “Improving black-box optimization in VAE latent space using decoder uncertainty” arxiv.org/abs/2107.00096 (1/)

Models reduce the need for data in decision-making but often pick up on spurious features. With Xiang, Ge, and Tommi Jaakkola, we propose TIA that relies on co-operative reconstruction for identifying important features in our #ICML2021 paper, xiangfu.co/tia

New #RSS2021 paper on self-supervised affordance discovery! GIFT: Generalizable Interaction-aware Functional Tool Affordances without Labels arxiv.org/abs/2106.14973 pair.toronto.edu/blog/2021/gift… w\ D. Turpin, L. Wang, Stavros Tsogkas, Sven Dickinson Vector Institute University of Toronto Robotics Institute

We need more papers like this. Simple and thought provoking on the foundations. I love the direction of performing adaptive computation at test time, and the analogy with human thinking. Should all models be recurrent?

Super excited to share something we have been working on for the last 1.5yrs. Check out our #RSS2021 paper on Rapid Motor Adaptation. RMA allows a legged robot trained *fully* in simulation to *adapt* online to diverse real-world terrains in real-time! ashish-kmr.github.io/rma-legged-rob…

Our new paper on "The Bayesian Learning Rule" is now on arXiv, where we provide a common learning-principle behind a variety of learning algorithms (optimization, deep learning, and graphical models). arxiv.org/abs/2107.04562 Guess what, the principle is Bayesian. A very long🧵

Are off-policy methods more sample efficient than on-policy in MARL? We show that PPO + centralized value functions is a strong baseline on a variety of cooperative MARL tasks! Paper: arxiv.org/abs/2103.01955 Blog post: bair.berkeley.edu/blog/2021/07/1…

Want to accelerate convex QP solvers using RL? Check out our latest work with Berkeley AI Research 👇 #ML4OPT #osqp

This paper: arxiv.org/pdf/2403.11940… is making an important point related to the tutorial (bit.ly/agentstate ) that Alex and I did at ICML. You really need forward dynamics on top of a multistep inverse model to fully capture an agent's endogenous dynamics.