DELE-w0.5: Inferring Action from Future Latent State for Robotic Manipulation
技术动态
来源：arXiv cs.RO发布时间待核实
arXiv:2608.22067v4 Announce Type: replace Abstract: World-Action Models (WAMs) build robot control on video-generation backbones, which jointly predict dense future visual trajectories and robot actions. We argue that video generation is an unnecessary intermediate objective for world-action modeling.