1 paper · 1 filter
Zhenyang Ni, Yijiang Li, Ruochen Jiao +7
Video generation models trained on heterogeneous data with likelihood-surrogate objectives can produce visually plausible rollouts that violate physical constraints in embodied man…