Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization
Zero-WAM uses a human demonstration video as the task prompt for a robot policy, without retraining.
The paper frames manipulation generalization as in-context learning: show the model how the task should unfold, then have it act. To train it, the authors build HumanGen, 74.2K paired human-robot examples spanning 8.6K tasks, generated from semantically matched robot trajectories and human videos. In RoboTwin 2.0 simulation, Zero-WAM reports a 47.0% average success rate on seven unseen tasks, 29.5 points above the strongest video-action baseline. Real-world tests are described as covering unseen configurations with multi-object scenes, long-horizon manipulation, and fine-grained insertion. HF Daily Papers' note
The paper frames manipulation generalization as in-context learning: show the model how the task should unfold, then have it act. To train it, the authors build HumanGen, 74.2K paired human-robot examples spanning 8.6K tasks, generated from semantically matched robot trajectories and human videos. In RoboTwin 2.0 simulation, Zero-WAM reports a 47.0% average success rate on seven unseen tasks, 29.5 points above the strongest video-action baseline. Real-world tests are described as covering unseen configurations with multi-object scenes, long-horizon manipulation, and fine-grained insertion. HF Daily Papers' note
score 5