Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Seeing Beyond the Next Step: World-Model-Guided Human-Like Navigation in Multi-Agent Scenes
SEONGEUN HONG, JuYeong Hwang, Jinhyun Kim, Hanyoung Jang, Hyeongyeop Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a world-model-guided navigation approach that evaluates candidate next actions by predicting their consequences in populated multi-agent scenes, to produce more human-like motion.
Problem
Choosing a next motion in populated scenes requires anticipating what each candidate action will cause, since failures such as drifting onto the road, cutting into pedestrian streams, or bottlenecking passages are invisible from the current state alone.
Approach
Uses a world model to roll out the consequences of candidate actions so the planner can reason about downstream effects on other agents and the scene before acting, rather than only avoiding immediate obstacles.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove both world models from preference refinement (ablation) | Ours · 5 held-out scenes avg | GSR / SRVR: GSR 77.4 -> 37.3; SRVR 4.0 -> 19.5 | Hurt |
| add state transition world model (ablation) | Ours · ETH_Intersection | GSR: 51.9 -> 83.3 | Improved |
| add semantic transition world model (ablation) | Ours · ETH_Hallway | SRVR: 13.7 -> 0.9 | Improved |
| use mixed top-5 + random-5 candidates instead of top-10 only (ablation) | Ours · Zara01_Group | GSR: 14.8 -> 77.8 | Improved |
| use mixed top-5 + random-5 candidates instead of random-10 only (ablation) | Ours · ETH_Intersection | GSR: 44.4 -> 83.3 | Improved |
| use pretrained VLM backbone instead of task-trained Transformer/ViT (ablation) | Ours · Zara01 | GSR: 48.1 -> 72.2 | Improved |
| distill world-model preferences into policy instead of running world models online | Ours vs OnlineWM · 5 held-out scenes avg | GSR: 80.6 (OnlineWM) vs 77.4 | No clear effect |
| distilled policy vs online planner inference latency | Ours vs OnlineWM · per decision | latency: 1.2 s -> 65 ms | Improved |
| use world-model-guided policy instead of trajectory predictors with MPC | Ours · 5 held-out scenes avg | GSR: 56.1 (Constant Velocity) -> 77.4 | Improved |
| use world-model-guided policy instead of GBM baseline on group distance | Ours · Zara01_Group | GD: 1.9 (GBM) vs 2.1 | Hurt |
| replace baselines with world-model-guided policy | Ours · Zara01_Group | SRVR: 17.4 -> 1.2 | Improved |
Closest papers by meaning
- NavWM: A Unified Navigation World Model for Foresight-Driven Planning (ECCV 2026 · similarity 0.91)
- Navigation World Models (CVPR 2025 · similarity 0.89)
- World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems (NeurIPS 2026 · similarity 0.88)
- WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.88)
- From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.87)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
