IMPACT: Attention Is the Interaction Map for Scalable Interaction-Aware World Model Training
Research
Source: arXiv cs.ROPublish time unverified
arXiv:2609.00161v1 Announce Type: cross Abstract: World models have made remarkable progress in action-conditioned future prediction for embodied agents, yet still struggle to model physically plausible interactions. Existing approaches address this limitation by constraining the generation process with external representations encoding motion, geometry, or semantics.