1 paper
Enyi Wang, Mingxin Wang, Quan Shi +12
World action models improve robotic manipulation by conditioning actions on predicted futures, yet existing tactile variants largely inherit video-generation pipelines that reconst…