具身智能观察

IMPACT: Attention Is the Interaction Map for Scalable Interaction-Aware World Model Training

技术动态

来源:arXiv cs.RO发布时间待核实

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.

IMPACT: Attention Is the Interaction Map for Scalable Interaction-Aware World Model Training | 具身智能观察