arXiv Robotics — research abstracts· Chen Zhao, Zipeng Ye, Yihan Lin, Yuchen Xie, Jing Zhang·· 16 hours agoEditorial score65
What Limits Cross-Task Generalization in World Action Models?
What Limits Cross-Task Generalization in World Action Models?
Summary
This research explores the limitations of cross-task generalization in world action models (WAMs) by separating them into world models and inverse-dynamics models (IDMs). The authors introduce the Align Once, Evaluate Everywhere (AoE) method, which allows testing across different future sources. They find that actions recovered from recorded futures perform better than those predicted by world models, and propose the VITA loss to improve execution success rates in generated futures.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org