MomWorld: Momentum-Aware Latent World Model for Long-Horizon Autonomous Driving
Abstract Overview
MomWorld is a latent world model for long-horizon autonomous driving that explicitly models both scene configuration and a latent momentum state. The method initializes these states from historical-to-current observations, rolls them forward with scene-adaptive retention, update, and reset mechanisms, and organizes the resulting future predictions into a Future World Memory for candidate plan scoring. It also introduces MoFlow, a momentum-conditioned flow-matching module that refines a selected base trajectory with bounded, horizon-aware corrections so near-term stability is preserved while longer-horizon adjustments remain possible. The paper evaluates the approach on nuScenes, NAVSIM v1/v2, Bench2Drive, and additional robustness settings.
Novelty
The paper’s main distinction is extending momentum from an ego-planning stabilizer into an explicit latent world-state variable that is jointly propagated with future scene state for planning. It further combines this with a momentum-conditioned flow-refinement stage that applies bounded, horizon-aware residual corrections rather than directly replacing the planned trajectory.
Results
On 6-second nuScenes planning, MomWorld reports the best average L2 error (1.17 m) and average collision rate (0.79%) in the compared table, improving over MomAD’s 1.42 m and 0.90%, and it reduces average collision rate by 12.2% relative to MomAD across the 1–6 s horizon. It also achieves the best reported long-horizon trajectory consistency on nuScenes (TPC average 1.19 vs. 1.42 for MomAD) and strong NAVSIM results, including 90.2 PDMS on v1 navtest, 90.1 EPDMS on v2 navtest, and 42.8 EPDMS on v2 navhard.
Key Points
- MomWorld jointly rolls out latent scene state and latent momentum, using maintain-update-reset dynamics to preserve stable motion trends while suppressing stale ones under scene changes.
- The Future World Memory produced from rollout is used for candidate trajectory scoring, and MoFlow refines the selected plan with momentum-conditioned residual flow plus horizon-aware bounded fusion.
- Empirically, the method improves long-horizon accuracy, consistency, and safety across nuScenes and NAVSIM, with ablations indicating that both the momentum-world modeling and the controlled flow-refinement components contribute to the gains.