Reproduce “Discrete-WAM: Unified Discrete Vision-Action Token Editing for World-Policy Learning” 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
Ziyang Yao, Haochen Liu, Yuncheng Jiang, Zeyu Zhu, Zibin Guo, Jingru Wang, Tianle Liu, Jianwei Cui and 5 more
Researchers at Xiaomi EV developed Discrete-WAM, a unified framework that employs discrete vision-action token editing for world-policy learning in autonomous driving. The system achieved an EPDMS score of 90.4 on the NAVSIM-v2 benchmark, outperforming existing state-of-the-art methods, and demonstrated high-fidelity world generation with an FID of 6.6 and FVD of 80.0.
Key insights
- Unifying visual observations, actions, decisions, and future states into a shared discrete token space enables stronger multimodal alignment and explicit compositional semantics compared to entangled continuous representations.
- Jointly training world models and policies within a single framework and optimizing shared token-level objectives effectively bridges the gap between understanding future dynamics and generating policy-relevant actions.
- Hierarchical policy construction, which predicts high-level decisions before refining low-level actions via parallel discrete diffusion and confidence-based token editing, provides an efficient and structured mechanism for generating temporally consistent long-horizon behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.