1 paper
Hao Wu, Tianyu Shao, Shengju Qian +10
Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons. Retaining the entire generated history does not necessarily im…