WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction
WAVE-0 adds a small latent world model so a robot policy can use predicted action consequences while choosing actions.
The paper says the 257M-parameter world model sits on top of a pretrained VLA and turns latent action plans into predicted future representations. Those predictions condition the action expert directly, without pixel reconstruction or large-scale video pretraining. The authors report 91.47% success on RoboTwin 2.0, 98.7% on LIBERO, and a real-world gain for π₀.₅ from 38.3% to 73.3%. Source: HF Daily Papers' note.
The paper says the 257M-parameter world model sits on top of a pretrained VLA and turns latent action plans into predicted future representations. Those predictions condition the action expert directly, without pixel reconstruction or large-scale video pretraining. The authors report 91.47% success on RoboTwin 2.0, 98.7% on LIBERO, and a real-world gain for π₀.₅ from 38.3% to 73.3%. Source: HF Daily Papers' note.
score 4