acceptodds
Under review as a conference paper at ICLR 2027

MomWorld: Momentum-Aware Latent World Model for Long-Horizon Autonomous Driving

Abstract

Long-horizon planning enables autonomous vehicles to anticipate scene evolution and potential risks, supporting safe and stable decisions in complex interactions. However, existing methods struggle to propagate motion trends from observed history into the future. Long rollouts based on a single latent state may further attenuate useful dynamics, retain stale motion patterns, and disrupt reliable near-term plans. We introduce MomWorld, a momentum-aware latent world model for long-horizon planning. MomWorld extracts scene motion trends from historical-to-current observations and propagates latent momentum into future horizons, jointly predicting future configuration and momentum states. A learnable momentum persistence mechanism preserves stable trends, scene-conditioned momentum updates adapt future dynamics, and a scene-adaptive reset gate suppresses stale momentum under abrupt changes. We further propose MoFlow, a momentum-conditioned flow-matching module that refines a base trajectory to align with the predicted future scene evolution in only a few integration steps, with a horizon-aware residual fusion that preserves near-term planning stability while permitting stronger long-range corrections. Extensive experiments on NAVSIM, nuScenes and Bench2Drive demonstrate that MomWorld improves long-horizon planning consistency and reduces the average collision rate by 12.2% relative to MomAD over a 6-second planning horizon.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Latent World Model, Autonomous Driving, Planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NBnuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MomWorld proposes a latent world model for temporally consistent, long-horizon autonomous driving. It jointly propagates scene states and a history-conditioned momentum variable through gated updates. The resulting future memory supports candidate trajectory scoring and selection. A flow-matching module then refines the selected trajectory through bounded, horizon-aware residual prediction. The paper reports improvements on nuScenes, NAVSIM and Bench2Drive.

Critical strengths
  • The paper's approach of combining historical dynamics with future scene prediction is a reasonable direction for improving planning consistency.
  • The future representations are used in both trajectory selection and refinement, giving them a clear role in the planning pipeline. It aslo includes several component ablations and reports the latency of different refinement settings.
Critical weaknesses
  • The motivation for the proposed modules could be better supported. The momentum variable appears similar to a learned recurrent hidden state, and its physical interpretation is not clear to me. Flow matching is a valid way to refine the deterministically selected trajectory. However, if the refinement process is also deterministic, it is unclear what benefit flow matching offers over a simpler residual regressor. A direct comparison would help clarify this choice.
  • The ablations show that these modules help, but without comparisons to simpler temporal and refinement models, I find it not eay to assess the benefit of the proposed design.
Questions for authors
  • The role of the momentum variable is not fully clear to me. Is the momentum given a physical definition? Have you compared the momentum dynamics with a GRU or another recurrent model of similar capacity, using the same inputs and supervision?
  • How does MoFlow compare with a simply residual regression model using the same conditioning and residual fusion?
  • Could you clarify the individual contributions of FWM, LWR, and SMD? The cumulative ablation shows gains as modules are added, but it is unclear how much each contributes and how their interactions affect the results.
  • Is GuideFlow's result with a scorer comparable under your NAVHARD evaluation protocol?
  • Could you point to the Bench2Drive results and appendix, and clarify how the six-second nuScenes baseline results were obtained?
2 comments

Thank you for the clarifications and additional experiments. The matched GRU comparison supports the structured state update, though I still view its methodological novelty as limited. For MoFlow, the gains over no refinement are modest, and the comparisons do not yet clearly distinguish the benefit of flow matching from iterative refinement more generally. While these designs have merit, their contributions and demonstrated benefits are not sufficient to change my overall assessment. I therefore will maintain my original score.

AuthorsOpenReview ↗

How MomWorld Uses the Predicted Scene Trend for Planning

Thank you for recognizing that the comparison with the matched GRU supports our structured state update. We would like to clarify one point: why this design matters for planning. During training, MomWorld uses differences between successive future scene representations to supervise a latent variable that captures the trend in scene evolution. At inference, the model uses its predicted trend to advance the scene state step by step. The resulting rollout serves both to select a candidate trajectory and to refine the selected trajectory. The matched GRU receives the same supervision and is integrated into the same planning pipeline, but its scene state at a given step need not be updated using the trend predicted at that step. Our design requires the model not only to predict how the scene will change, but also to use that prediction to construct the future scene representations on which planning relies.

Reviewer_FRHaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a novel latent world modeling paradigm with momentum awareness. The momentum can adaptively retain useful history trajectories and deny stale motions. This is achieved by simultaneously predicting both latent scene configurations and momentum states. Another momentum-induced flow matching module with residual fusion is also designed to strengthen the near-term consistency and long-term corrections. Experiments are extensive, where abundant results on NAVSIM, nuScenes, and Bench2Drive prove the state-of-the-art performance of MomWorld.

Critical strengths
  1. The paper has a clear motivation: this paper focuses on dealing with long-horizon predictions, which are influenced by accumulated errors, sudden state variations, and uncertainty. To address these problems, the authors design both momentum-based world states for memory retrieval and also momentum-enhanced flow matching for near-term consistency.

  2. Overall, the paper seems to be novel in this field. Although MomAD proposes a similar momentum-based autonomous driving system, while they fail to model the future imaginations.

  3. Experiments are extensive. Authors demonstrate the strengths on three benchmarks including NAVSIM, nuScenes, and Bench2Drive. These experiments can prove the effectiveness of the proposed contributions.

  4. The paper is easy to follow.

Critical weaknesses
  1. In the Scene-Adaptive Momentum Dynamics, the authors design three gating modules to predict retention, reset, and momentum parameters. However, the detailed network architecture is not given. More detailed explanations are required for better understanding how these modules are designed.

  2. In the future world memory, could authors change the position of momentum and latent states? In other words, the momentum state is kept and latent state is used for residual. Since the network predicts the momentum and latent states at the same time, I am wondering the differences between both. Would they have similar functions in the network but be given different meanings only?

  3. In Formula (8), the scale is fixed as 50. How is this parameter decided? Also, in Formula (14), how are the weights in residual fusion determined? More descriptions and also analysis are needed.

  4. The paper lacks comparisons with other concurrent works such as DriveVA (ECCV'26).

Questions for authors
  1. Figure 4 only have one sample with three timestamps, which are hard to explain the main claim in long-horizon predictions. Longer prediction results and more visualization samples are needed. Otherwise, the main strength is somewhat overclaimed.

  2. I would like to see how the momentum is visualized and functioned in the network. Only planning trajectories cannot show how the designed momentum state really works.

  3. Efficiency is also a crucial metric in real-world planning, I am curious about the latency time comparison with other works.

Reviewer_kpUVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MomWorld, a momentum-aware latent world model for long-horizon autonomous driving. The method maintains separate latent scene and momentum states and jointly propagates them into the future using a gated update mechanism. The predicted states are organized into Future World Memory to support candidate trajectory selection. The paper also introduces MoFlow, which uses momentum-conditioned flow matching and horizon-aware residual fusion to refine the selected trajectory. Experiments on nuScenes, NAVSIM, and Bench2Drive show improvements in planning accuracy, temporal consistency, and safety.

Critical strengths
  • The idea is interesting by explicitly modeling motion trends alongside scene states for long-horizon planning, especially when traffic dynamics change over time.
  • The paper evaluates on several benchmarks, including open-loop, pseudo-simulation, and closed-loop settings. The overall results are good.
  • The paper is well-structured.
Critical weaknesses
  • The proposed momentum update is quite similar to conventional gated recurrent updates, particularly GRU. Propagating dynamic memory through latent world models has already been explored in works such as DriveWorld. The novelty of the momentum-aware world model is thus not sufficiently presented.
  • A key motivation is to preserve useful motion trends and suppress old momentum during long rollouts. However, the experiments mainly evaluate the final planning performance. They do not directly show whether the predicted scene states and momentum are more accurate, especially over longer horizons or during sudden changes.
  • The method uses three losses for future query prediction, momentum supervision, and auxiliary future-state prediction. However, there is no ablation showing how much each loss contributes. In addition, the fixed loss weights and the manually defined horizon-aware fusion schedule lack sensitivity analysis, for example, 0.5 and 0.2, 0.05.
  • The writing can be further improved. The motivation for using separate retention and reset gates is not sufficiently discussed, especially considering their similarity to existing gated recurrent mechanisms. The relationships between different modules, the relationships between the different losses and the corresponding model outputs could also be explained more clearly.
Questions for authors
  • The latent world rollout is not conditioned on candidate ego trajectories, so all planning candidates share the same predicted future representation. This may limit the model's ability to capture action-dependent interactions, particularly in scenarios where other agents react differently to different ego decisions.
  • The paper argues that jointly propagating scene and momentum states helps preserve useful dynamics and suppress outdated motion trends. However, it is unclear why these benefits cannot be achieved with a conventional recurrent latent state that already captures temporal information. The current ablations mainly evaluate different components within the proposed architecture, rather than the necessity of the scene and momentum decoupling.
  • The paper describes momentum as a representation of motion trends, but its supervision is based on differences between consecutive scene features. Such differences may also reflect viewpoint changes, occlusion, or changes in scene appearance, rather than motion alone.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.