ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training
ZimaBlue reports a zero-shot robot success jump from 36.1% to 77.8% when trained with more than 120,000 hours of embodied video.
The paper frames egocentric human and robot video as a scalable substitute for scarce action-labeled robot trajectories. Its training pipeline pre-trains on action-free embodied video, grounds that model with heterogeneous robot trajectories, then specializes it to a target robot. A Slow-Fast architecture is used so the larger world model can support a lightweight branch that predicts actions at 30 Hz on an NVIDIA RTX 4090. The authors say the gains are strongest on unseen tasks. HF Daily Papers' note
The paper frames egocentric human and robot video as a scalable substitute for scarce action-labeled robot trajectories. Its training pipeline pre-trains on action-free embodied video, grounds that model with heterogeneous robot trajectories, then specializes it to a target robot. A Slow-Fast architecture is used so the larger world model can support a lightweight branch that predicts actions at 30 Hz on an NVIDIA RTX 4090. The authors say the gains are strongest on unseen tasks. HF Daily Papers' note
score 6