안녕하세요, 데이터사이언스메이드 토토사이트원에서 구글 딥마인드 연구자들을 모시고 BK21 x ERC 세미나를 개최하오니 많은 관심과 참여를 부탁드립니다.
일시: 2025. 9. 24(수) 오전 11:00 - 오후 12:00
장소: 43동 101호
Speakers: Alex Bewley, Krzysztof Choromanski, David Rendleman (all from Google DeepMind)
Title: 3D Robotics Transformers
Abstract: We will present some of the recent 3D Transformer models built by the Google DeepMind Robotics Team. The talk will focus on models leveraging a new class of relative positional encoding methods, known as STRING (ICML 2025) (project site), generalizing the celebrated RoPE mechanism, and particularly relevant for 3D modeling (with tokens embedded in 3D+ dimensional metric spaces). We will show how STRING can robustify 3D robotic policies by making them to generalize better to different conditions (e.g. lighting). We will also present High-Density Visual Particle Dynamics (HD-VPD) (CoRL 2024) point cloud based world models, leveraging a new Transformer architecture, called Interlacer (using both Performers and GNNs), and capable of modeling point clouds of sizes 100K+ (with tokens corresponding to individual points). We will conclude the talk with the linear-attention Transformers, applying the recently introduced mechanism of Graph Random Features (GRFs) (ICML 2023). Those linear-attention Transformers with GRFs (ICLR 2025) can be used in particular to further enhance the aforementioned HD-VPD models or in general for efficient graph data modeling.
Based on collaboration with Mithun Jacob, Connor Schenck, Joshua Ainslie, Isabel Leal, Rene Wagner (all Google DeepMind), and many others