Seeing Beyond the Next Step

World-Model-Guided Human-Like Navigation in Multi-Agent Scenes

Seong-Eun Hong1* JuYeong Hwang1* Jinhyun Kim1 Hanyoung Jang2 HyeongYeop Kang1†
1Korea University2NC AI

*Equal contribution   †Corresponding author

NeurIPS 2026

Paper (coming soon) arXiv (coming soon) Code (coming soon)
vs
Ours
time→

Method

World models are used only at training time, as a preference oracle. At deployment they are removed and the policy acts in a single forward pass.

Overview of the two-stage training pipeline

Dual World Models

State Transitionwho moves where
Semantic Transitionwhat the scene affords
Input images, ground-truth segmentation, and Semantic Transition World Model predictions over time

Scenes

Students003, Hotel, Zara01, ETH, ETH_Hallway and ETH_Intersection rebuilt in Unity