Reproduce “Map-World: Masked Action planning and Path-Integral World Model for 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
Bin Hu, Zijian Lu, Haicheng Liao, Chengran Yuan, Bin Rao, Yongkang Li, Guofa Li, Zhiyong Cui and 2 more
MAP-World introduces a prior-free, anchor-free, and reinforcement learning-free multi-modal planning framework for autonomous driving, utilizing masked action planning and a path-integral world model. The system achieved a state-of-the-art Predictive Driver Model Score (PDMS) of 88.8 on the NAVSIM test set and maintains a real-time inference latency of 13.5 ms.
Key insights
- The Masked Autoencoder (MAE) paradigm can be extended to forecast future states within a latent world model for autonomous driving, enabling prior-free multi-modal trajectory generation.
- Leveraging a path-integral formulation allows the world model to learn from an expectation over all generated multi-modal future possibilities, rather than just a single selected one, for more robust foresight.
- Achieving genuinely multi-modal, diverse, and natural motion planning for autonomous vehicles is possible without restrictive anchors or cumbersome RL-based selection mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.