acceptodds
Under review as a conference paper at ICLR 2027

DriveWeave-Policy: Weaving Sparse Future Imagination and Trajectory Planning through Aligned Denoising

Abstract

A central challenge for driving world-action models (WAMs) is how to effectively leverage visual generative priors for trajectory planning. Yet driving scenes are highly dynamic, and densely generating future visual observation imposes a substantial modeling burden on trajectory planning. Separate diffusion branches offer a straightforward way to reuse pretrained generative models in driving WAMs. However, coordinating these branches through LLM-mediated feature exchange alone may be insufficient to ensure consistency between imagined futures and planned actions. In fact, imagination and planning reflect the same underlying scene dynamics, and we accordingly introduce DriveWeave. Unlike existing frame-by-frame autoregressive unified WAMs, DriveWeave is designed to better harness visual generative priors at a lower sampling frequency. Within this framework, we compare three training paradigms for learning and transferring cross-modal priors: Imagine-then-Plan, Plan-then-Imagine, and Coupled Denoising. We find that sparse anchored imagination and coupled denoising best match the tight perception–action loop of driving. Furthermore, we introduce a learnable trajectory encoding method that brings the distributions of low-dimensional trajectory features and high-dimensional visual features closer together in a shared latent space. We coordinate visual and trajectory denoising through two complementary mechanisms: cross-modal information exchange and dynamic alignment. For information exchange, shared attention enables bidirectional conditioning at every denoising step. For dynamic alignment, we reformulate the reverse-diffusion dynamics with a consistency-driven coupling term that adjusts modality-specific updates based on agreement between predicted futures. On NAVSIM, DriveWeave achieves a PDMS of 91.3, the best performance among the evaluated methods.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, World-Action Models, Diffusion Models, Trajectory Planning, Multimodal Generation

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_7PWFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how sparse future visual prediction can support autonomous-driving trajectory planning. DriveWeave jointly denoises trajectory, auxiliary future ego-state, and multi-view visual latents using a pretrained video DiT. A trajectory codec aligns latent distributions, shared attention exchanges information, and a pose-consistency energy is applied during both training and sampling. The default model predicts one terminal visual anchor while retaining all eight waypoints over four seconds. On the original NAVSIM navtest protocol, it reports 91.3 PDMS. Ablations separate information exchange, consistency training, sampling correction, and visual-target density.

Critical strengths
  1. The paper provides a concrete way to coordinate future visual prediction and trajectory planning. Shared attention exchanges information, while a common ego-pose representation allows the model to compare what the predicted trajectory and imagined observations imply about motion. Applying this consistency objective during both training and sampling is a sensible design. Table 4 supports the distinction: consistency training improves PDMS from 87.9 to 90.0 without sampling correction, and the full method reaches 91.3. This shows that the benefit extends beyond an additional inference-time adjustment.

  2. Sparse future imagination offers a useful accuracy–efficiency trade-off. The model retains all eight waypoints over a four-second horizon while predicting only a terminal multi-view visual anchor. In Table 3, this achieves 91.3 PDMS, compared with 85.1 for plan-only prediction and 90.1 for dense visual prediction; the reported latency falls from 340.2 ms for dense prediction to 66.2 ms. These results provide useful evidence that planning can benefit from a compact future visual target without paying the cost of a dense visual rollout.

Critical weaknesses
  1. The evidence for visual–trajectory consistency relies heavily on the same learned pose readout used to enforce that consistency. The frozen readout R_eta converts generated visual latents into an ego-pose estimate, and both the consistency loss and the reported E_pose use this estimate (Secs. 3.3 and 4.1). A lower E_pose therefore establishes agreement under the training proxy, but does not independently verify that the generated views depict the motion specified by the trajectory. The PDMS improvements remain meaningful; what is less well established is the visual-consistency explanation for them. The paper refers to a future-image shuffling test in Appendix D, but its results and the readout's held-out pose accuracy are not available in the supplied PDF. Validation of this central component is needed to support the claimed geometric coordination.

  2. The contribution of trajectory–visual distribution alignment is not isolated experimentally. ActionConvAE combines trajectory reconstruction with an MMD objective, and the paper argues that bringing trajectory latents closer to visual latents helps reuse the pretrained video backbone (Sec. 3.2). However, Fig. 4(a) shows marginal density profiles, while the planning ablations in Tables 2–4 keep the codec fixed. These results do not establish whether MMD-based alignment improves planning over a reconstruction-only codec, or whether the gain comes simply from using a learned trajectory representation. A controlled codec comparison, reporting both waypoint reconstruction accuracy and downstream PDMS, is needed to substantiate this part of the method.

Questions for authors
  1. How accurately does the frozen visual-motion readout recover ego pose on held-out recorded image pairs, and what does the future-image shuffling test show? For generated views, does the reduction in E_pose also correspond to better visual–trajectory agreement under an independent motion estimate? Please provide the relevant Appendix D results and distinguish readout accuracy from agreement among the model's own predictions.

  2. How does ActionConvAE with MMD compare with the same codec trained only for reconstruction, keeping the DiT training and inference settings fixed? Please report waypoint position/heading reconstruction errors and downstream PDMS. This would clarify whether matching the visual latent distribution provides a planning benefit beyond the learned trajectory encoding itself.

2 comments
AuthorsOpenReview ↗

Response to 7PWF: Pose Consistency, MMD Alignment, and Generation Validity

Thank you for your recognition of our paper,your valuable comments greatly improved its quality.

W1/Q1: Table S3 obtains 0.47 m/2.8° on held-out real image pairs; shuffling future images raises the error to 2.36 m/11.7°, indicating the readout uses future-image motion information and pose accuracy is guaranteed. Your concern may be robustness. However, Ev itself is generated and contains errors far larger than the readout error, so a slightly worse readout has little effect. The first consistency is pose-consistency(E_pose) term, acting on the sampling-denoising formula, optimizing both the denoising velocity field and the added consistency cost, so the model understands multimodal denoising. We also use this as a simple evaluation metric to assess the validity of generation. Existing generative backbones still fit single-modality generation; for multimodality they treat visual and action tokens as the same kind tokens for denoising. We want more than per-token imitating GT: even with low MSE loss, one cannot tell which tokens are good, and action results may still be poor. We internalize autonomous-driving requirements—unified heading and observation angles—into the reverse dynamics of denoising, adding constraints between tokens for more effective multimodal generation. Ev-GT and Ea-GT correctness is achieved by optimizing MSE loss during training so they learn correct results. Pose-consistency guarantees valid generation, not correct generation. Table S4 uses a visual pose estimator independent of training/sampling guidance to evaluate generated images' 4-s endpoint position; independent evaluation also shows improvement relative to GT.

W2/Q2-Revision: Added a simpler tiling results row to Table S5. Table S5 already isolates the MMD increment: keeping ActionConvAE and the reconstruction objective but removing MMD drops PDMS from 91.3 to 90.2 and raises the independently estimated visual–trajectory endpoint difference from 2.35 m to 2.55 m, supporting “adding distribution matching on this codec is beneficial.” The fixed-tiling row shows a sharp drop, clearly demonstrating this distribution matching’s role. The second 'consistency' is alignment(MMD) term, acting on information-interaction understanding. This is a problem we found during training: in the early stage, trajectory generation was already good, but image quality still did not meet requirements. This indicates trajectory denoising did not well use visual-generation priors; the two were largely generated separately. Alignment makes information interaction between denoising directions and velocities of trajectories and images more effective, enabling better exploit information from both. Lifting the eight-segment polyline to the same feature distribution as images lets the model learn what it should look like as observations change from current to future. Although supervising directly 4s ahead is difficult, it is necessary: a long trajectory helps the world model better understand evolution rules, whereas small segment-by-segment supervision may merely imitate. Pose-consistency(E_pose) ensures valid generation in sampling; alignment(MMD) ensures sufficient information interaction to learn correct generation in training.

Thank you for the detailed response. The held-out pose results and the shuffling test address my questions about what the readout learns, and I appreciate the independent evaluation on generated views. It would help to report the error against the ground-truth endpoint separately from the gap between the visually estimated and planned endpoints, since the latter is the agreement I was asking about.

The comparison with and without MMD addresses my main concern about the codec. The PDMS drop from 91.3 to 90.2 with the same ActionConvAE provides evidence that MMD contributes beyond the learned trajectory representation. Reporting position and heading reconstruction errors for both variants would make the comparison more complete. Overall, the response addresses most of my concerns, and I am keeping my rating.

Reviewer_UkCUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper argues that a good drive world model should jointly estimate the trajectory and observations. Therefore, the paper proposes three training recipes for joint estimation, introduces trajectory priors learned by a customized autoencoder to enhance video generation, and incorporates pose information into the generation process. Finally, to alleviate the heavy training burden for dense generation, it uses asynchronous generation for trajectory and video frames. Results show that the method achives better results than prior approaches.

Critical strengths
  • The method demonstrates that involving pose information in video generation enhances performance in the driving world model task.
Critical weaknesses
  1. Unfair comparisons

While Table 1 and 3 shows that the proposed method performs better, the gains are not fully verified due to the misalignment in setups.

1.1 Multi-view v.s. Single-view

Both Epona and DriveWAM reconstruct one view in one time step, but the proposed method reconstructs 3 views simultaneously ("Sensors" in Table 1). It's not clear whether the performance gain stems from the proposed method or the stronger supervision provided by more complete scene coverage.

1.2 Dense image sampling

In Table 3, the paper compares the proposed Sparse anchor approach to a Dense target. However, the "Dense frame-by-frame" target is not a typical setup used in other approaches. For instance, DriveWAM performs video generation under 1Hz sampling (see Sec 4.2 in DriveWAM), yielding 4 frames with a 4-second time window, which is quite similar to the proposed approach, which generates 3 frames within a 4-second window.

Therefore, the proposed sparse anchors seem to address the exact multi-view generation implementation used in this paper. However, the multi-view generation itself is not comprehensively discussed.

  1. Inaccurate claims

The paper claims that "These WAMs typically use multi-branch designs that denoise modalities separately, limiting cross-modal interaction to LLM-mediated feature exchange without explicitly coordinating generative update" (L48). However, not all methods follow the scheme that the paper argues against. For instance, DriveWAM already performs trajectory and video generation within one Transformer, with the LLM features acting only as a condition. The cross-modal interactions are feasible in DriveWAM. The paper can benefit from better positioning its contributions more precisely relative to previous methods.

  1. Diagrams are hard to read

While I appreciate the comprehensive statement in "AI USE STATEMENT", the diagrams are hard to interpret and understand in that:

  • some text is in extremely small fonts that are hard to read (eg, Fig 2, Fig 4.b)
  • some text and legends are too blurry to recognize (eg, Fig 5)
Questions for authors

According to the Weaknesses section, I suggest the authors:

  • add the results of a single-view variant to the comparison in Table 1
  • discuss the necessity of incorporating a multi-view reconstruction target in this approach
  • revise the claims or clarify if there is a misunderstanding

Minor issue: L39-40, left quotation mark in LaTeX

"what the scene will look like" --> ``what the scene will look like''

2 comments
AuthorsOpenReview ↗

Response to UkCU: Clarifications on Fair Comparisons, Core Motivation, and Presentation

Thank you very much for your suggestions. We can see you read our paper carefully; your valuable comments greatly improved its quality.

  1. Regarding unfair comparisons:

1.1-Revision: Added single-view variant in Table 1. Previous end-to-end/VLA work recognizes multiple views are necessary. Theoretically they provide stronger supervision, but effective use in autonomous-driving WAM is difficult. Existing methods are not unwilling to try multi-camera setups (datasets are open-source); their compatibility with the base model is poor, as explained later. Under fully fair conditions, our method also performs well. Thus both the method and more complete scene coverage bring improvements, both part of our work.

1.2 Your comments inspired us; we should discuss multiple-views utilization/generation more, a manifestation of the visual-prior upper bound. Multi-view sparse object anchoring improves efficiency in using effective visual information. “Frame-by-frame” was used to compared with high-sampling-frequency approaches: predicting three views every four seconds vs. front view every second means only one fewer image for optimization-not much difference. However, we supervise different views over a longer time span, whereas they supervise the same view over a shorter time span. We mean we achieve better results under more difficult supervision, with higher visual-prior utilization efficiency.

2 Our core motivation is the abstract's first sentence: maximally use visual-generation priors for trajectory planning. Yet some design choices overemphasized our thinking, so readers may mistake that paradigm is our actual work; we will weaken this narrative and better emphasize true joint denoising. We start from the perspective of the base model, aiming to train a backbone that truly performs joint denoising of trajectories and images, with autonomous driving optimization to address multi-view inputs, complex scene changes, and inference speed. Existing WAM work mostly reuses original video-generation models. Their consideration is mainly to make their inputs better fit base models, rather than modifying base models to fit the domain tasks. In fact, the base DiT cannot well handle joint understanding/generation of trajectories and multi-view images; what it excels at is single-view images and text input. The introduced multi-branch denoising and joint denoising with VLM reasoning only show they avoid the essential problem: although DriveWAM treats VLM results only as inputs, VLM is used to convert trajectory-image understanding into rich text to fit the base model. Epona similarly treats DiT only as a module, multi-branch denoising is used because the base model adapts poorly to multimodal generation. Although metrics improve, without base-model optimization, problems arise: fragmented generated targets, inability to optimize jointly, inference latency, and costly text annotation. In fact, the original generative base model's optimization direction/results are not well suited to autonomous driving, especially after introducing trajectories to build a WAM; information flow/interaction inside the future-denoising Transformer is also unclear. Our work explores the base model itself: information interaction during training, and optimization of the denoising principle during sampling and generation. If interested in details, can refer to e8Xh's W2/Q2. Returning to multi-view utilization in 1.1, without such fundamental experimentation/research, adapting on the original video-generation base model is very difficult. We also start from an application perspective—autonomous driving's inference-speed requirements and text-annotation cost—to conduct such supervision and report inference time, not just metric improvements.

3 — Revision: All Figures We apologize for the poor reading experience caused by our oversight: during revision the original images were too large, making fonts extremely small. The three right-side images in visualization are generated and resolution-adapted, so clarity is limited. But it shows some straight-driving and turning choices are more conservative than human driving (row 2 considers the left white car starting; row 3 turns more inward), indicating the model learned real autonomous-driving rules from denoising information and image changes of deceleration/stopping trajectories across scenes, rather than merely imitating ground truth. Regarding the difficult supervision in 1.2, we believe it is necessary: a long trajectory segment helps the world model better understand evolution rules, whereas small segment-by-segment supervision may merely imitate. It is our most exciting result: without overusing text, WAM truly combines trajectory and image information and realizes multi-vehicle interaction (e.g., the white car possibly changing lanes), not simply using a video-generation model for ego-trajectory generation. This is the starting point of building a world model, not just chasing metrics.

I truly appreciate the author's endeavor in providing additional results and explanations.

Regarding Weakness 3, my concern has been addressed: The authors provided a revised figure, which is clearer to read.

For Weakness 2, the concern has been mostly addressed, as the authors highlight the difference between the proposed approach and prior works. However, the revised claim does not match the statement in the paper perfectly, which may require further revision.

For Weakness 1, thanks to the authors for providing new results, and they look clearer now. However, I found that the single-view version (DriveWeave Coupled- SV in Table 1) performs comparably to DriveWAM in most metrics, which indicates that the performance gain mainly comes from additional output images instead of the contributions/techniques illustrated in the paper. To highlight the real source of improvement (the introduction of multi-view generation for driving tasks), the paper needs substantial revision.

Therefore, although I appreciate the authors' additional results, the effort required to revise the paper seems to go beyond a rebuttal phase. I therefore lean toward keeping my score of weak reject.

Reviewer_e8XhOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates how generated future scenes can support planning, particularly how to align a world model’s imagined futures with future actions. To address inconsistencies between predicted images and trajectories, as well as the potentially unnecessary modeling burden of dense future-image prediction, the authors generate sparse future keyframes and map the motion implied by images and trajectories into a common space constrained by a consistency energy. By coordinating future generation and planning, the approach aims to improve overall autonomous-driving performance.

Critical strengths
  1. The framework is thoughtfully developed. Through ActionConvAE, shared denoising, pose-consistency checks, a consistency energy, and sparse future anchors, it systematically explores internal consistency in driving world models. Beyond allowing future-image and trajectory predictions to exchange information, it uses discrepancies in a common pose space to enable mutual correction during generation.
Critical weaknesses
  1. The paper’s central motivation is not presented clearly. Although it primarily examines how inconsistencies between future generation and planned actions may degrade performance, the exposition is overly verbose. The abstract and introduction do not lead naturally to this central problem, leaving the main idea obscured.
  2. The proposed solution appears overly complex. Several mechanisms are introduced to promote consistency between future generation and planning, but their responsibilities seem to overlap, giving the impression that they repeatedly address the same problem. This makes their distinct roles difficult to understand.
Questions for authors
  1. Could the authors revise the presentation of the central problem, particularly in the abstract and introduction, to make the motivation and main argument more explicit? This would help readers understand the problem the framework is intended to address.

  2. Could the authors concisely explain the distinct role of each module in addressing inconsistencies between future generation and planning? Are these mechanisms complementary, or could they introduce conflicting effects? For example, could generating fewer future images weaken consistency? An explanation supported by the relevant evidence would help assess whether the complexity of the framework is justified.

  3. Given the many possible variations in the position and orientation represented by future images, can the authors clarify whether a precise distance metric can uniquely characterize pose consistency? This would help assess the interpretation and adequacy of the proposed consistency measure.

1 comment
AuthorsOpenReview ↗

Rebuttal for e8xh: Clarifications on Novelty, and Presentation

Thank you for your suggestions, which greatly help improve our paper's quality and readability.

W1/Q1-Revision: Abstract, Introduction. Our core motivation is the abstract's first sentence: how to maximally use visual-generation priors for trajectory planning. Yet we then introduced a question that diverted readers to this point. We only meant to use autonomous-driving characteristics to qualify this motivation's difficulty, but did not expect misunderstanding. Some design choices overemphasized our thinking and caused redundancy; we will improve narrative. We hope not waste your time or effort and our explanation can give you different thoughts and a deeper understanding of the field—our honor.

We actually aim to raise the upper bound of denoising generative models in using trajectories and observations. Existing WAM work, including DriveWAM, Epona..., mostly reuses original video-generation models. Their consideration is mainly make their inputs better fit base models, rather than modifying base models to fit the domain tasks. In fact, the base excels at single-view images and text input. Therefore, we conduct fundamental research to train a backbone that truly performs joint denoising of trajectories and images, with autonomous driving optimization to address multi-view inputs, complex scene changes, and inference speed. The introduced multi-branch denoising and joint denoising with VLM reasoning only show that they avoid the essential problem: Multi-branch denoising is used because base model adapts poorly to multimodal generation; VLM is used to convert trajectory-image understanding into rich text to fit the base model. Although metrics improve, without optimizing the base model, various problems will arise, such as fragmented generated targets, inability to optimize jointly, inference latency, and costly text annotation. We may have overemphasized surface problems and weakened our main idea.

W2/Q2: Our work explores the base model itself: information interaction during training, and optimization of the denoising principle during sampling and generation. The first pose-consistency(E_pose) mainly term acts on the sampling-denoising formula, making it optimize not only the denoising velocity field but also the added consistency cost. Existing generative backbones still fit single-modality generation; using them for multimodality will treat visual and action tokens as the same kind of token for denoising. We want more than each token imitating ground truth: even with low MSE loss, one can't tell which tokens are good, and action result may still be poor. We internalize autonomous-driving requirements for generated trajectories and images—unified heading and observation angles—into the reverse dynamics of denoising, adding extra constraints between tokens for more effective multimodal generation. The second alignment(MMD) term acts on information-interaction understanding. This is a problem we found during training: in the early stage, trajectory generation was already good, but image quality still did not meet requirements. This indicates trajectory denoising did not well use visual-generation priors; the two were largely generated separately. Alignment makes information interaction between denoising directions and velocities of trajectories and images more effective, enabling better exploit information from both. Pose-consistency ensures valid generation in sampling; alignment ensures sufficient information interaction to learn correct generation in training. Regarding information interaction, we further adjust the model with a slot mask to control conditions and outputs, and use three training schemes to clarify information flow and mutual trajectory-image utilization in joint denoising. We reveal accurate future information as supervision to test whether each use it well. Results show that under separate denoising, trajectories better exploit image priors, while the trajectory's 24 numbers have little effect on image optimization. Under joint denoising, what early-learned priors interact with is still noise from the other modality. under pose-consistency generation, Prior instead harms overall denoising. It takes more time matching our full model performance—showing the importance of synchronous information interaction. For "fewer future images weaken consistency?" and multi-view sparse anchoring, you can see UkCU 1.1 and 1.2. If interested our final capabilities differ from the original base model, see the visualization analysis in UkCU 3, which metrics can't reflect.

Q3: Our measure does not aim to uniquely characterize a future image or the complete future scene. It compares planar ego poses—position and heading—at matched timestamps in the current ego coordinate frame. We also distinguish consistency from correctness. The visual readout is learned and itself can make errors. This is also detailed in Tables S3/4.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to 31LV: Pose Consistency, Joint Denoising, MMD Alignment, and Additional Evaluations

Q1: Table S3 obtains 0.47 m/2.8° on held-out real image pairs; shuffling future images raises the error to 2.36 m/11.7°, indicating the readout uses future-image motion information and pose accuracy is guaranteed. Your concern may be robustness. However, Ev itself is generated and contains errors far larger than the readout error, so a slightly worse readout has little effect. The pose-consistency(E_pose) term mainly acts on the sampling-denoising formula, optimizing both the denoising velocity field and the added consistency cost, so the model understands multimodal denoising. Existing generative backbones still fit single-modality generation; for multimodality they treat visual and action tokens as the same kind tokens for denoising. We want more than per-token imitating GT: even with low MSE loss, one cannot tell which tokens are good, and action results may still be poor. We internalize autonomous-driving requirements—unified heading and observation angles—into the reverse dynamics of denoising, adding constraints between tokens for more effective multimodal generation. Pose-consistency guarantees valid generation, not correct generation. Table S4 uses a visual pose estimator independent of training/sampling guidance to evaluate generated images' 4-s endpoint position; independent evaluation also shows improvement relative to GT.

Q2-Revision: Added table S7 The main purpose is to clarify information flow and mutual trajectory–image utilization during joint denoising, not to compare performance gaps. Because existing work treats DiT only as a module, information flow/interaction inside the future-denoising Transformer is unclear, especially after introducing trajectories to build a WAM. Since supervision differs, we only tested whether each can use accurate future information revealed in advance. Results show that under separate denoising, trajectories better exploit image priors, while the trajectory's 24 numbers have little effect on image optimization. Under joint denoising, early-learned priors still interact with noise from the other modality; under pose-consistency generation, the prior instead harms overall denoising. Matching our full model takes more time, showing the importance of synchronous information interaction. 10K steps already converge trajectories, and 30K is more than enough for images; to align training cost, we adopted these two configurations. We only want to show that although our supervision target is more difficult, training difficulty has not changed significantly.

Q3-Revision: Added a simpler tiling results row to Table S5. Table S5 already isolates the MMD increment: keeping ActionConvAE and the reconstruction objective but removing MMD drops PDMS from 91.3 to 90.2 and raises the independently estimated visual–trajectory endpoint difference from 2.35 m to 2.55 m, supporting “adding distribution matching on this codec is beneficial.” The fixed-tiling row shows a sharp drop, clearly demonstrating this distribution matching’s role.

Q4-Revision: Added NAVSIM v2 evaluation table S6. The final visual anchor via pose-consistency mainly ensures the validity of final multimodal generation. Intermediate consistency is better understood through alignment (MMD)-based information interaction during denoising, which is a black-box process. Lifting the eight-segment polyline to the same feature distribution as images lets the model learn what it should look like as observations change from current to future. Although supervising directly 4 s ahead is difficult, it is necessary: a long trajectory helps the world model better understand evolution rules, whereas small segment-by-segment supervision may merely imitate. This consistency only ensures, at the token level during denoising, synchronization of denoising velocity/direction and consistency of their respective evolution.

Q5: Latency spans from encoding multi-view images/trajectory inputs to trajectory decoding. It includes pose readout and gradient computation for sampling correction, but excludes future RGB decoding. Main results are run with seeds 195/196/197 and averaged; we will clarify this. I understand your concern: metric gains are small, but some models’ effects are not well reflected by metrics alone. Visualizations show some straight/turning choices are more conservative than human driving (row 2 considers the left white car starting; row 3 turns more inward), indicating the model learned real autonomous-driving rules from denoising information and image changes of deceleration/stopping trajectories across scenes, not mere ground-truth imitation. It is our most exciting result: without overusing text, WAM combines long trajectory and image information, helping the world model better understand evolution rules, not simply using a video-generation model for ego-trajectory generation.

AuthorsOpenReview ↗

Thank You to All Reviewers

Thank you all reviewers for your time and effort. Your comments have played a very important role in systematically organizing the quality of my paper, and have also helped me understand this field more deeply and systematically. Previously, perhaps only I could understand it; now I feel it is much clearer. Thank you again for the guidance of all the professional reviewers. My reply may be a little late, but I did not want to let down your time and effort; my responses also reflect my in-depth understanding and thoughts about this field. I noticed that the confidence levels are not very high. Some ideas and inferences about black-box issues may be treated by AI as directly observed facts, since this is my colloquial narration, but if these can help you understand this field more easily, it doesn't matter.

Thank you for the detailed responses and additional experiments. These help clarify the method, but some key concerns remain insufficiently addressed. Considering the paper’s novelty, technical soundness, and experimental evidence, I believe my original rating still reflects its strengths and limitations. I will therefore maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.