Grounded Action Model Bets on 3D Grounding as the Foundation for Robotics, Fully Open-Sourced

joemeno · x · 2026-10-13

• Researchers from UW, Northwestern, and NUS introduced the Grounded Action Model (GAM), questioning whether VLAs and world models are the right foundations for robotics and proposing "ground first, then act."
• Pipeline: language or a 2D point selects task objects; a frozen promptable 3D detector (WildDet3D, built on SAM 3) supplies image features, 3D boxes, size, orientation, and point clouds; a flow-matching transformer predicts actions.
• All checkpoints, training code, and evaluation code for RoboTwin, LIBERO-PRO, and real-world experiments are open-sourced on GitHub (MIT license).

Original post →

More from Embodied

• Proposing a 'Millennium Problem of Robotics': ten long-horizon research challenges listed — DJiafei · 2026-10-13
• Open-source wearable motorless leader arm captures teleoperation data for robot learning — chris_j_paxton · 2026-10-13
• Apple launches HomePod mini 2 with S10 chip and Sound Glow at $149 — Scobleizer · 2026-10-13
• Natura-1 zero-shot demo runs on Wuji Hand 2 in one uncut 30-min take — chris_j_paxton · 2026-10-13
• Dev Hooks ChatGPT Realtime Voice Model Into an Apple Watch — halo-w3fsd32 · 2026-10-13
• From-scratch robot GHOST001 shows first limb jumps in R002 update — wavefnx · 2026-10-13

Embodied channel →