Reworks world-action models for robot control so they predict sparse visual deltas plus actions instead of dense future frames, decoupling action-conditioned dynamics from nuisance appearance changes. Adds Streaming Delta Memory (SDM) to update cached anchor context with compact deltas instead of reprocessing full observations through the heavy video expert. On RoboTwin, improves average success over Fast-WAM from 81.3% to 85.4% (clean) and 75.8% to 83.9% (visual randomization), cuts training FLOPs 17.8-23.8%, and cuts one-step inference latency/FLOPs by ~36.6%/31.6%; also leads on real-world evaluations. Code and website released.
