Uni-World VLA: Interleaved World Modeling and Planning for Autonomous Driving

Qiqi Liu^1,2,3*, Huan Xu³*, Jingyu Li^1,2,3, Bin Sun³†, Zhihui Hao³†, Dangen She³, Xiatian Zhu⁴, Li Zhang^1,2‡

¹Fudan University; ²Shanghai Innovation Institute; ³Li Auto Inc.; ⁴University of Surrey
* equal contribution; † project leader; ‡ corresponding author

---

📰 News

---

🔭 Project Overview

---

💡 Key Features

• 🔄 Interleaved world modeling and planning: alternates future frame prediction and ego action/trajectory generation step-by-step, forming a closed-loop interaction that keeps planning conditioned on imagined observations.
• 🤖 Unified autoregressive VLA formulation: generates visual tokens and action queries in a single sequence, tightly coupling prediction and control under temporal causality.
• 📹 Depth integration for geometric cues: augments historical frames with monocular depth maps and fuses geometry features via cross-attention to improve long-horizon scene prediction.

---

📊 Results

Table 1. Closed-loop planning results on NAVSIM.

Table 2. World modeling / prediction results on NAVSIM.

Visualization.

---

🧾 TODO

• Release arXiv paper
• Release code
• Release model weights

---

📖 Citation

@article{liu2026uniworld,
  title   = {Uni-World VLA: Interleaved World Modeling and Planning for Autonomous Driving},
  author  = {Liu, Qiqi and Xu, Huan and Li, Jingyu and Sun, Bin and Hao, Zhihui and She, Dangen and Zhu, Xiatian and Zhang, Li},
  journal = {arXiv preprint arXiv:2603.27287},
  year    = {2026},
}