Robotics & WM Reading Club 30: What Scale Doesn't Fix: Representation & Architecture in Robot Manipulation+Omni-Reason Video Model. SF 9/26

Robotics & World Models Reading Club 30: What Scale Does Not Fix: Representation and Architecture in Robot Manipulation. Berkeley 9/26 A high-signal reading group for AI researchers & builders pushing the frontiers of robotic world models, WAMs, and embodied intelligence. In our previous sessions, we brought together researchers and engineers from Boston Dynamics, Google DeepMind, NVIDIA, Stanford, UC Berkeley, Physical Intelligence, Tesla, Generalist, Rhoda AI, and leading Bay Area robotics startups. Hosted by Junfan Zhu & Aurora Feng. Support Saturday Robotics Inc: https://donate.stripe.com/28EcN52rjgeY1fJboYgEg00 ​​​​Reading Club 30's Core Theme What Scale Does Not Fix: Representation and Architecture in Robot Manipulation Keynote: Devol Robots (Elijah Yi Herng Ong - CTO) Robot foundation models have scaled rapidly over the past two years, yet scale alone has not produced reliable physical interaction. This talk presents two architectures from Devol Robots that attack the problem from opposite ends: what a model represents, and how that model is organized. The first is the Interaction World Model (IWM 1.0), which addresses representation. Built on Interaction Geometry, a Riemannian framework that keeps robot sensorimotor signals on the manifolds where they physically live rather than flattening them into Euclidean vectors, IWM infers compliance from the shape of a demonstration itself and commands the robot through an impedance controller whose stiffness and damping the model predicts, with no force or torque signals anywhere in training. On contact-rich tasks from socket insertion to sub-millimeter lens seating it reaches 92% average success against 38% and 50% for leading VLA and world model baselines, using roughly an order of magnitude fewer action parameters. The second is Devol-ONE, which addresses architecture. Existing world-action models keep prediction and policy structurally separate and connect them only through a predicted output. Devol-ONE instead runs vision-language reasoning, latent world prediction, and action generation as three streams of a single autoregressive Mixture of Transformers that exchange information at every layer, so imagination shapes execution continuously rather than at one bottleneck. It reaches 98.4% on LIBERO and 71.4% zero-shot on LIBERO-Plus without training on the perturbed data, and holds its performance on RoboTwin 2.0 under environment randomization while a strong pixel-space baseline collapses from 76.6% to 2.7%. Elijah will show how these two lines converge: a representation that captures the physics of contact, carried by an architecture that reasons about the future while it acts. Together they define what a deployable and reliable manipulation system looks like in industrial settings, and he will demonstrate what each looks like running on production hardware today. ​​​​​​​​​​​​​Pre-Readings www.devolrobots.ai/whitepaper ​​​​​​​​​​​Location Berkeley ​​​​​​​​​​​​​​Date & Time Saturday, September 26, 2026 | 2:00 PM – 5:00 PM ​​​​​​​​​​​​​​​Join Discord Community https://discord.gg/WH7DrTHRXK ​​​​​​​​​​​​​​Follow Saturday Robotics on X /saturdayrobotic ​​​​​​​​​​​​​​Agenda 2:00 PM – 2:30 PM Door Opens & Social Food 😋, beverages🧋 and UNLIMITED strawberries 🍓 (our official reading club fruits ☺️😄). 2:30 PM – 3:30 PM Keynote by Devol Robots YouTube Recording: TBD (We are looking for recording volunteers) 3:30 PM – 5:00 PM Q&A, ​open-floor roundtable (10–20 min per topic) on spotlight papers or any paper you’d like to highlight. Feel free to share why the paper matters and its technical details. ​Future events #iros-reading-club-31-0928: 🍾 IROS 2026 x Saturday Robotics — Robotics Research Night | Reading Club 31. Pittsburgh 9/28 Session 31 Luma: https://luma.com/tzbw7n61 ​​​​​​​​​​​​​​​​​​​​Past events #reading-club-28-0912: Booster T2: The Next Frontier of Open Humanoid Robotics Session 28 Luma: https://luma.com/00z3oxw6 #reading-club-27-0905: Rethinking Robot Development: Co-Designing Morphology, Sensing, and Learning Chenyang Ma (Applied Intuition, UNC) Session 27 Luma: https://luma.com/jm9jceqr #reading-club-26-0829: Sunnyvale 8/29 Session 26 Luma: https://luma.com/enowqo5v #reading-club-25-0822: Contact-Rich Robot Learning from Human Videos and Tactile. SF 8/22 Kelin Yu (Maryland, Amazon FAR) Session 25 Luma: https://luma.com/yl76re1b #reading-club-24-0820: Saturday Robotics x EAI @ WRC Beijing | PHYSICAL AI UNPLUGGED —— 与 Physical AI 研究者的一场非正式夜谈 • 李晓帆 — Head of World Models, X Square Robot (自变量机器人) • Tongzhou Mu — Research Scientist, Rhoda AI • Haoyi Niu — University of California, Berkeley • Yufei Wang — Member of Technical Staff, Genesis AI • Shibo Zhao — Research Scientist, Amazon FAR • 郑思鹏 — Technical Partner, Being Beyond • Tianle Zhang — Embodied AI Researcher, JD Explore Academy • Yihang Li — Embodied AI Researcher, JD Explore Academy Session 24 Luma: https://luma.com/jezexfza #reading-club-23-0815: Engineering Robotic Simulators for Evaluation and Beyond. SF 8/15 Kaifeng Zhang (Columbia, World Labs) Session 23 Luma: https://luma.com/non72bev #reading-club-22-0808: ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow. SF 8/8 Haoyi Niu (UC Berkeley) Session 22 Luma: https://luma.com/7nnl4a5r Reading Club 22 Review: https://x.com/junfanzhu98/status/2086482247344079100?s=20 Reading Club 22 Recap: https://x.com/junfanzhu98/status/2086484540013560280?s=20 #reading-club-21-0801: Vision-Language-Kinematics Supervision for Perception-Based Humanoid Loco-Manipulation — SF 8/1 Yen-Jen Wang (UC Berkeley, Amazon FAR) Session 21 Luma: https://luma.com/4xkibxbh Reading Club 21 Review: https://x.com/junfanzhu98/status/2083948325112533456?s=20 Reading Club 21 Recap: https://x.com/junfanzhu98/status/2083950393734975839?s=20 #reading-club-20-0725: Agentic Robotics Models (ENPIRE and Cap-X), Mountain View 07/25 Haoru Xue (UC Berkeley) Kris Hauser (Samsung Research America, Robot Intelligence Lab) Session 20 Luma: https://luma.com/