What it is
A world action model for robot manipulation that fuses video dynamics, VLM semantics, and 4D priors, pretrained on 23K hours of open heterogeneous data. Unlike WAMs that need future-video rollout or weaker visual priors, it couples a pretrained video expert and action expert in a directed MoT. Causal Imprint turns future supervision into action-relevant predictive features, while 4D priors are added by training-only distillation.
Key results
- 92.8% on LIBERO-Plus, best under distribution shift
- 71.9% on RoboTwin 2.0 Clean2Random; 90.0% on Clean2Clean
- 23.9% on RoboDojo, nearly 2x the strongest prior WAM
Predictive video supervision helps robot control most when converted into action-facing features rather than requiring test-time future generation.