Megadose The AI curve, compressed.

ResWAM: Learning What Changes, Not What Stays, for World-Action Models

· HF Daily Papers ·
ResWAM reports 80% average success on real-robot Open, Put and Stack tasks, 10 percentage points above π₀.₅, at 92.7 ms latency.

The authors argue that predicting entire future observations wastes learning on static backgrounds. Their method instead supervises changes in frozen DINOv3 features, concentrating on the robot arm and manipulated objects. The 2.3-billion-parameter model jointly learns actions and visual changes through a lightweight dual-branch architecture. Without external robot-action pretraining, it also reports top average success rates on LIBERO-Plus, RoboTwin 2.0 and RoboDojo, with the lowest training and inference compute among the compared models.

HF Daily Papers' note

score 5

Categories: Research