Abstract
Autonomous driving requires navigating complex environments where multiple plausible future trajectories exist. Traditional end-to-end systems often rely on predefined trajectory anchors, which can be computationally expensive or lack the flexibility to handle diverse road geometries. This paper introduces Map-World, a prior-free planning framework that utilizes Masked Action Planning (MAP) to treat future motion as a sequence completion task. Combined with a Path-Integral World Model, Map-World learns from a full distribution of potential futures rather than a single "best" path. The proposed method achieves state-of-the-art results on the NAVSIM benchmark with a PDMS score of 88.8 and maintains real-time inference speeds.
Introduction
A critical challenge in autonomous driving is handling multi-modality—the fact that multiple safe paths may exist at any given moment (e.g., turning left vs. going straight). Most current models address this by selecting from a fixed library of "anchors." However, anchors present a trade-off: a small set may fail in complex geometries, while a large set increases computational overhead. Furthermore, many systems discard alternative safe trajectories during training, losing valuable information. Map-World addresses these issues by replacing handcrafted anchors with a generative approach that understands the entire spectrum of possible outcomes.
Methods
Map-World introduces two primary innovations: Masked Action Planning and a Path-Integral World Model.
Masked Action Planning (MAP)
Inspired by Masked Autoencoders, MAP treats trajectory planning as a sequence completion task. The model processes past states and a coarse driving intention, then uses a Transformer decoder to reconstruct "masked" future waypoints. By injecting noise into a compact latent state, the system generates diverse and temporally consistent trajectories without requiring a predefined anchor library.
Path-Integral World Model
Rather than evaluating a single chosen path, Map-World employs a lightweight world model to simulate future Bird's Eye View (BEV) semantics for all candidate trajectories. Using a differentiable path-weighting scheme inspired by Feynman’s path-integral formulation, the training loss is calculated as an expectation over all generated modes. This allows the planner to learn from the entire distribution of plausible futures in a fully differentiable manner.
Experiments and Results
The framework was evaluated on the NAVSIM and nuScenes datasets.
Performance on Benchmarks
Map-World sets a new state-of-the-art on the NAVSIM benchmark, achieving a PDMS score of 88.8, outperforming existing models like DiffusionDrive. On the nuScenes validation set, it achieved a remarkably low average L2 error of 0.50m and a collision rate of 0.07%, demonstrating superior safety and precision compared to anchor-based methods.
Efficiency and Qualitative Analysis
Despite its multi-modal nature, Map-World is highly efficient, with an inference latency of 13.5 ms on an NVIDIA 4090. Qualitative results show that the model generates trajectories that align more closely with ground-truth human driving, particularly in complex intersections where baseline models often struggle.
Conclusion
Map-World demonstrates that a prior-free, generative approach to trajectory planning can outperform traditional anchor-based systems. By leveraging masked sequence completion and path-integral weighting, the model achieves robust, safe, and efficient autonomous driving. Future work will explore scene-adaptive mode allocation to dynamically adjust the number of predicted paths based on environmental complexity.
Summary updated