Reproduce “Drive-HWM: Hierarchical World Models for Dynamic-Latent Guided Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoxin Fan, Tianbao Zhang, Wenjun Wu, Xiaofeng Wang, Yeying Jin, Jian Zhao, Zheng Zhu, Shuicheng Yan
Researchers from Beihang University, Shanghai Jiao Tong University, National University of Singapore, and others developed Drive-HWM, a hierarchical slow-fast world model for autonomous driving that integrates long-horizon future prediction with immediate action generation. This approach achieved the highest overall scores on NAVSIM v1 (PDMS 93.8) and NAVSIM v2 (EPDMS 86.4) benchmarks, while reducing inference latency by 28% compared to a single-world-model baseline.
Key insights
- Explicitly separating long-horizon future representation prediction from high-frequency action generation allows each component to operate optimally at its respective temporal scale.
- Supervising the slow world model with optical flow effectively creates "Dynamic-Aware Latents" that encode motion dynamics, providing more pertinent future context for driving decisions than other visual targets.
- Using Feature-wise Linear Modulation (FiLM) allows for adaptive injection of dynamics-aware future predictions into a pre-trained multimodal backbone, conditioning action generation without disrupting its internal representation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.