GreenWAM: A World Foundation Model for Robotics
GreenWAM is pitched as one robotics model that can predict worlds and act across many robot bodies.
The paper says its shared action space covers 20 robot families, putting different machines into common joint coordinates. GreenWAM is trained in visual, action-conditioned, and world-action modes, then adapted into policies by separating an action expert from the generator. The authors report top-tier world-model results on WorldArena 1.0 and strong policy results on RoboTwin 2.0, LIBERO-plus, CALVIN, and LIBERO. They also adapt it to a real humanoid and say it outperforms policies adapted from pi_0.5 and Cosmos 3 Edge in that setup. HF Daily Papers' note
The paper says its shared action space covers 20 robot families, putting different machines into common joint coordinates. GreenWAM is trained in visual, action-conditioned, and world-action modes, then adapted into policies by separating an action expert from the generator. The authors report top-tier world-model results on WorldArena 1.0 and strong policy results on RoboTwin 2.0, LIBERO-plus, CALVIN, and LIBERO. They also adapt it to a real humanoid and say it outperforms policies adapted from pi_0.5 and Cosmos 3 Edge in that setup. HF Daily Papers' note
score 5