Adding future-frame DINOv3 CLS alignment to the action stream of VAE-based world-action models improves out-of-distribution success without hurting in-distribution performance.
Motus: A unified latent action world model
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models
Adding future-frame DINOv3 CLS alignment to the action stream of VAE-based world-action models improves out-of-distribution success without hurting in-distribution performance.