acceptodds
Withdrawn from ICLR 2027

ReDrive: Shaping Representations with World Modeling for End-to-End Driving

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Driving policies require capabilities of scene understanding and future evolution prediction. To achieve this goal, current end-to-end models typically construct complex perception-planning pipelines or introduce world models that explicitly predict future states, resulting in a complex system architecture. Inspired by the transferability of general-purpose visual representations, we argue that combining sufficiently strong visual representations with representation world modeling can support effective planning without relying on complex inference-time auxiliary modules. Based on this insight, we present ReDrive, an end-to-end driving framework that strengthens planning-oriented visual features via future representation prediction. To achieve this, ReDrive adopts a three-stage training pipeline consisting of driving video pretraining, joint world-modeling and planning training, and planner adaptation. This yields a strong planning-oriented representation and a high-performance planner, while requiring neither auxiliary perception modules nor future prediction at inference time. Experiments on NAVSIM demonstrate strong performance, achieving 91.0 PDMS on NAVSIM v1 and 90.8 EPDMS on NAVSIM v2. These results show that shaping representations with world modeling is sufficient to enable high-performance end-to-end planning while retaining a simple encoder-planner inference pipeline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Driving, Representation Learning, Predictive Learning, Trajectory 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_UgrqOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ReDrive, which uses future prediction to supervise the representations learned by a visual encoder. Compared with conventional future-prediction approaches, the method does not require generating future video frames during inference. Instead, it distills the representational capabilities associated with future generation into the visual encoder during training. The method uses vjpeca2 for pretraining on autonomous driving scenes, followed by multistage training. It achieves strong results on several mainstream benchmarks, demonstrating the feasibility of retaining future-prediction capabilities within the visual encoder.

Critical strengths

The paper offers useful insights to the community: strong performance can be achieved by using future prediction for video representation learning without generating future video frames during inference. The ablation studies clearly show that future prediction itself provides only modest gains, whereas improvements in visual representations yield substantially larger gains. The ablation studies are clear and easy to follow, and the comparative conclusions are well articulated.

Critical weaknesses

The paper analyzes the future predictor’s behavior solely through the similarity of predicted features under different action conditions. This demonstrates sensitivity to actions but is insufficient to establish that the predictor can provide reliable planning gradients for non-expert trajectories.

Questions for authors

Could the authors provide a quantitative comparison between predicted features and actual future features, together with an interpretable analysis of the future predictions? This evidence would help assess prediction reliability beyond sensitivity to action conditions and inform the evaluation of whether the predictor can provide reliable planning gradients for non-expert trajectories.

3 comments

We sincerely thank the reviewer for the positive evaluation of our work and the valuable suggestions. We have conducted additional experiments to address the remaining concerns and provide detailed responses below.

Question 1:How reliable and interpretable are the future representations predicted by the future predictor?

Response 1: We thank the reviewer for the suggestion. To evaluate the reliability of the predicted future representations, we conduct additional experiments on 1,000 NAVSIM test samples. We compare the predicted future representations conditioned on GT and randomly shuffled trajectories against the actual future representations. As shown in Table R1, GT trajectory conditioning achieves higher cosine similarity, indicating that the future predictor produces representations more consistent with the actual future under the observed trajectory condition.

Table R1. Similarity between predicted and actual future representations under different trajectory conditions.

Trajectory ConditionCosine Similarity ↑
Shuffled GT Trajectory0.5063
GT Trajectory0.6184

We further evaluate prediction consistency through a trajectory ranking experiment. For each sample, we rank the GT trajectory and randomly sampled trajectories from other samples based on the cosine similarity between predicted and actual future representations. GT Top-1 measures the proportion of samples where the GT trajectory ranks first, while Pairwise Win Rate measures how often it outperforms shuffled trajectories. As shown in Table R2, the future predictor captures meaningful correspondence between trajectory conditions and actual future representations.

Table R2. Ranking accuracy of the GT trajectory among different trajectory candidates.

CandidatesGT Top-1 (%) ↑Pairwise Win Rate (%) ↑
891.698.67
1686.398.75
2482.298.81
4873.198.88

Question 2: Can the future predictor provide reliable planning gradients for non-expert trajectories?

Response 2: We thank the reviewer for this question. To evaluate gradient reliability on non-expert trajectories, we conduct a one-step gradient update experiment on 1,000 planner-generated trajectories. As shown in Table R3, gradient descent on the future prediction loss reduces the GT trajectory loss in 70.9% of samples, compared with 46.4% for random directions. These results support the effectiveness of the frozen future predictor in providing useful planning gradients.

Table R3. Effectiveness of future prediction gradients on planner-generated trajectories.

Update DirectionGT L1 Improvement Ratio ↑
Random Direction46.4%
Emb Gradient Descent70.9%

Thank you for the detailed response and additional experiments. The comparison with observed future features and the trajectory-ranking results address my concern about prediction reliability. The one-step experiment also provides direct evidence that the predictor can offer useful gradients for planner-generated trajectories. These results address my main concerns, and I maintain my weak accept recommendation.

We sincerely thank the reviewer for your positive feedback and for acknowledging our clarifications. We greatly appreciate your time and thoughtful consideration.

Reviewer_d4wtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReDrive, an end-to-end autonomous driving framework that improves planning-oriented visual representations through future representation prediction. It introduces a three-stage training pipeline consisting of driving-domain self-supervised pretraining, joint training of trajectory planning and future representation prediction, and planner adaptation using a frozen future predictor. During inference, only the visual encoder and planner are retained, eliminating the need for additional future prediction modules. Experiments on NAVSIM v1 and v2 demonstrate strong planning performance, including on the challenging NavHard benchmark.

Critical strengths
  1. The three-stage training pipeline. The paper presents an effective training strategy that progressively combines self-supervised video pretraining, joint representation learning and planning, and planner adaptation.
  2. Strong performance on NavSim. ReDrive achieves competitive results on both NAVSIM v1 and v2, including the challenging NavHard benchmark, demonstrating the effectiveness of the proposed approach.
Critical weaknesses
  1. Limited novelty. The proposed framework mainly combines existing techniques, including V-JEPA2-based representation learning, future feature prediction, and flow-matching-based planning. The technical novelty beyond these existing components is somewhat limited.
  2. Lack of closed-loop evaluation. Although the paper reports strong results on NAVSIM, it lacks evaluation in a fully interactive closed-loop environment. Such experiments would provide stronger evidence of the planner's robustness and practical effectiveness.
Questions for authors

See weaknesses.

1 comment

We sincerely thank Reviewer d4wt for your thoughtful review and for recognizing the strengths of our work, including the clarity of our training framework and its strong performance. Below, we address your concerns in detail.

Question 1: What are the key contributions and novelty of ReDrive?

Response 1: We thank the reviewer for the question. The key contribution of ReDrive lies in achieving strong planning performance by leveraging pretrained video representations without complex architectural designs or additional perception supervision. Existing methods that leverage pretrained visual representations often introduce additional task-specific designs, while world modeling approaches typically require complex architectures to incorporate future information into planning. In contrast, ReDrive introduces trajectory-conditioned future representation prediction to provide future-aware supervision for visual representation learning and planning optimization. This simple yet effective design achieves strong planning performance without additional architectural complexity.

Question 2: Can ReDrive be evaluated in interactive closed-loop environments?

Response 2: We thank the reviewer for this valuable suggestion. To address this concern, we conducted additional zero-shot closed-loop evaluations on two benchmarks, HUGSIM[1] and Neuro-NCAP[2]. As shown in Tables R1 and R2, ReDrive outperforms all compared methods on the reported overall metrics, demonstrating its effectiveness and generalization in closed-loop driving environments. We will include these results in the revised paper.

Table R1. Zero-shot closed-loop evaluation on HUGSIM.

MethodNC ↑DAC ↑TTC ↑COM ↑PDMS ↑RC ↑Easy ↑Medium ↑Hard ↑Extreme ↑Average ↑
VAD41.290.328.095.328.330.142.08.57.76.313.9
UniAD65.693.251.666.349.443.864.037.221.06.331.2
LTF44.392.837.594.836.538.066.115.512.011.723.1
DrivoR52.295.646.293.944.847.278.029.120.014.132.5
ReDrive69.294.962.576.057.457.473.555.031.317.644.3

Table R2. Zero-shot closed-loop evaluation on Neuro-NCAP. All methods are evaluated without trajectory post-processing.

MethodScore ↑Collision Rate (%) ↓
VAD0.6692.5
UniAD0.7388.6
SparseDrive0.9293.9
BridgeAD-B1.6072.6
VaVAM-L2.4657.9
ReDrive2.6252.9

[1] Zhou H, Lin L, Wang J, et al. Hugsim: A real-time, photo-realistic and closed-loop simulator for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

[2] Ljungbergh W, Tonderski A, Johnander J, et al. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. European Conference on Computer Vision, 2024.

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

The paper uses the representations learned by pretrained encoders, namely VJEPA, and extends them to downstream planning tasks. The authors propose a three-stage training recipe. First, they continue V-JEPA 2 pretraining on front-camera driving videos from nuScenes, nuPlan and NVIDIA-Physical AI datasets. Second, they jointly train a flow-matching DiT planner and a future-representation predictor on top of the encoder. The predictor is conditioned on the ground-truth trajectory, and its loss updates only the encoder and the predictor. In the last stage, they freeze the encoder and the predictor and train only the planner, using imitation losses together with the feature-prediction loss computed through the frozen predictor on the planner’s own trajectories. At inference only encoder and planner are used to generate trajectories. The resulting planner is evaluated on the NAVSIM v1 and NAVSIM v2 benchmarks.

Critical strengths
  • S1. The overall idea of adapting a pretrained video encoder to driving data and training a simple planner on top of it is interesting. At inference ReDrive needs only the encoder and the planner, without perception heads or future prediction.
  • S2. The authors ablate the contribution of each training stage. Table 7 also tests whether the improvement stays in the encoder by freezing it and training a new planner.
  • S3. The final PDMS achieves SOTA on both benchmarks.
Critical weaknesses
  • W1. The improvements appear to come mainly from the V-JEPA2 backbone rather than from world modeling. V-JEPA2 with a DiT planner already reaches 88.9 (Table 6). This is already above most of the models reported in table 1. The overall gain is insignificant. According to Table 6, the pretrained V-JEPA2 with a DiT planner achieves 88.9 PDMS. The adaptation to the driving domain increases this number by only 0.3, and the full three-stage training adds 1.3. The numbers are extremely close to each other and no variance or number of runs is reported. The frozen-encoder experiment should include an encoder trained only with the planning loss and results over several seeds are needed to attribute the gain to world modeling.

  • W2. Using latent future prediction as a training-only auxiliary objective is not new. LAW (ICLR 2025) predicts future latent features conditioned on the planner’s predicted trajectory to improve the planning representation. SSR (ICLR 2025) uses a future-feature predictor conditioned on the planned trajectory only during training. Latent-WAM (Wang et al., 2026a) in driving follows similar idea. The V-JEPA2 driving backbone is also shared with Drive-JEPA. The remaining contributions appear to be GT-trajectory conditioning (Stage 2) and Stage 3, which resembles LAW’s predicted-trajectory conditioning.

  • W3. The importance of Stage 3 is not justified. Stage 3 is presented as a main contribution, but it adds only 0.3 PDMS and appears in a single ablation. All other analyses stop after Stage 2, the paper does not state whether the headline result includes it, and its effect on NAVSIM v2 and navhard is not reported. Its gain is also not isolated, since Stage 3 adds imitation losses and extra training steps alongside the feature-prediction loss (Eq. 7). Moreover, the target of the feature-prediction loss is the real future, which belongs to the logged trajectory, so the loss is lowest when the planner reproduces the logged trajectory. It therefore at best acts as a second imitation signal and penalizes other valid trajectories, although the paper describes Stage 3 as on-policy adaptation. Also, in Section 3.3.2 says gradients from future prediction are kept out of the planner to avoid “steering the policy toward actions that are easier to predict.” Stage 3 then does exactly that. Please reconcile this, and discuss whether the planner can exploit the frozen predictor off-distribution.

  • W4. The advantage over generative world models is claimed but not tested. The paper argues that generative world models that predict future observations may be distracted by fine-grained details, so that predicting latent representations is the better choice. Epona is compared in Table 1, but as a complete independent system with a different backbone, planner and training, so the comparison does not isolate the representation. Epona (ICCV 2025) and Orbis (NeurIPS 2025) are action-conditioned driving world models trained on driving video. Using the representation of one of these models in place of the V-JEPA 2 encoder, with the same Stage 2 and Stage 3 training, would show whether the advantage comes from the proposed three-stage approach on latent representations.

  • W5. Some methods in the v2 table are missing from the v1 table. No inference latency or parameter count is reported and compared with existing works, although a simple inference pipeline is the main motivation.

Questions for authors
  • Q1 (W1). Could you add a Table 7 row with an encoder trained in Stage 2 with the planning loss only, then frozen with a new planner? Could you report the main results and Tables 5 to 7 over several seeds?

  • Q2 (W2). How does ReDrive differ from LAW and SSR? How does conditioning Stage 2 on the predicted trajectory, as LAW does, compare with conditioning on the ground truth?

  • Q3 (W3). Why are Tables 4, 5 and 7 and Fig. 4 reported only after Stage 2? Could you report them after Stage 3, and show the effect of Stage 3 on NAVSIM v2 and navhard?

  • Q4 (W3). Does the planner produce trajectories in Stage 3 on which the frozen predictor is unreliable? How does individual loss in stage three improve the model?

  • Q5 (W4). Could you use the representation of a generative driving world model, such as Epona (ICCV 2025) or Orbis (NeurIPS 2025), in place of the V-JEPA 2 encoder, with the same Stage 2 and Stage 3 training? This would test the claim of the paper and show whether the three-stage approach is advantageous.

  • Q6 (W5). Could you also report inference latency and parameter counts for ReDrive and the main baselines?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aTwP 1/3

We sincerely thank the reviewer for the constructive comments and valuable suggestions. We have conducted additional experiments and provided further clarifications to address the concerns. We respond to each question in detail below.

Question 1: Method improvements and multi-seed evaluation. (W1 & Q1)

Q1a: Evaluation of a planning-only encoder baseline under the frozen-encoder protocol in Table 7.

Response 1a: We thank the reviewer for this suggestion. We add a planning-only encoder baseline under the same evaluation protocol, where each encoder is frozen and evaluated with a newly initialized planner. As shown in Table R1, incorporating world modeling further improves the performance beyond planning supervision alone, demonstrating its contribution to visual representation learning.

Table R1. Evaluation of encoders obtained from different training strategies.

PretrainPlanning LossWM LossPDMS ↑
✓✗✗86.7
✓✓✗89.2
✓✓✓90.2

Q1b: Performance gains over the strong pretrained visual backbone.

Response 1b: ReDrive aims to leverage strong pretrained visual representations and further improve them through world modeling for downstream planning. Under the 4-frame encoder setting, ReDrive improves PDMS by 1.3 points over the V-JEPA2 + DiT baseline, closing approximately 22% of the performance gap to the human reference (94.8 PDMS). This represents a substantial improvement over an already strong baseline. Furthermore, as shown in Table R6, Stage 3 achieves larger performance gains on NAVSIM v2 and NavHard, particularly in more challenging driving scenarios.

The 0.3-point gain from driving-domain pretraining reflects the strong representations learned through large-scale pretraining. Subsequent planning training also adapts the encoder to driving data, further reducing the additional benefit of domain-specific pretraining.


Q1c: Multi-seed evaluation.

Response 1c: We thank the reviewer for the suggestion. Given the substantial computational cost, we conduct three-seed evaluations of the main results. As shown in Table R2, ReDrive achieves consistent performance across different random seeds.

Table R2. Multi-seed evaluation of ReDrive.

BenchmarkSeed 42Seed 0Seed 1
NAVSIM v1 (PDMS ↑)90.9691.0590.98
NAVSIM v2 (EPDMS ↑)90.8090.6690.57
NavHard (EPDMS ↑)34.3734.9834.78

Question 2: Method comparison and trajectory conditioning strategies. (W2 & Q2)

Q2a: Comparison with existing methods.

Response 2a: We thank the reviewer for the comments. ReDrive investigates how to obtain planning-oriented visual representations through a simple yet effective approach. We show that combining strong video self-supervised representations with future representation prediction enables high-performance driving without additional perception supervision or complex inference-time world models.

The key distinction of ReDrive lies in achieving strong planning performance by adapting general-purpose visual representations through world modeling on driving data and planning supervision alone, without complex architectural designs or additional supervision. This is a non-trivial achievement. LAW relies on task-specific visual representations, while SSR focuses on compact navigation-guided BEV representations with auxiliary future prediction. In contrast, ReDrive builds on strong pretrained video representations and explicitly uses future prediction to improve both representation learning and planning through staged optimization. Latent-WAM incorporates geometric knowledge distillation from external models to enhance spatial representations, whereas ReDrive requires neither geometric distillation nor perception supervision. Drive-JEPA also leverages pretrained video representations, but pretraining alone provides limited planning gains, requiring additional multimodal trajectory distillation and specialized planning designs. In contrast, ReDrive achieves strong performance by simply introducing future representation prediction.


Q2b: GT vs. predicted trajectory conditioning in Stage 2.

Response 2b: We thank the reviewer for this valuable suggestion. We evaluate a LAW-style variant using predicted trajectories with gradients propagated to the planner. As shown in Table R3, ground-truth trajectory conditioning achieves better planning performance, supporting our choice of GT conditioning for representation learning in Stage 2.

Table R3. Comparison of trajectory conditioning strategies in Stage 2 with a 4-frame encoder.

MethodNC ↑DAC ↑EP ↑C ↑TTC ↑PDMS ↑
Pred_Traj (LAW-style)98.894.281.5100.095.587.0
GT_Traj (ReDrive)99.197.283.5100.096.589.9

AuthorsOpenReview ↗

Response to Reviewer aTwP 2/3

Question 3: Stage 3 evaluation, loss ablation, and optimization strategy. (W3 & Q3–Q4)

Q3a: Why are Tables 4, 5, 7, and Fig. 4 evaluated only after Stage 2?

Response 3a: We thank the reviewer for this comment. The evaluation stage was selected based on the purpose of each experiment.

  1. Table 5 specifically investigates the effectiveness of Stage 2 training and is therefore evaluated at the end of Stage 2.
  2. Fig. 4 analyzes the future predictor and Table 7 evaluates the visual representations. Since the encoder and future predictor are frozen in Stage 3, their parameters remain identical to those in Stage 2, and the analysis results are unchanged. We will remove the Stage 2 wording in the revised paper to avoid confusion.
  3. For Table 4, we additionally report the Stage 3 results in Table R4.

Table R4. Performance of different visual backbones after Stage 2 and Stage 3.

BackboneStage 2 PDMS ↑Stage 3 PDMS ↑
DINOv284.484.9
MAE83.984.4
V-JEPA289.990.2

Q3b: Stage 3 loss ablation.

Response 3b: We thank the reviewer for the suggestion. We conduct a Stage 3 loss ablation under the same training budget. As shown in Table R5, removing either the trajectory loss or future prediction loss degrades planning performance, demonstrating that both losses contribute to Stage 3 and their combination achieves the best results.

Table R5. Ablation study of Stage 3 training objectives with a 16-frame encoder.

SettingFM LossTraj LossFeature LossPDMS ↑
Stage 2———90.8
Stage 3 w/o Traj Loss✓✓90.6
Stage 3 w/o Feature Loss✓✓90.7
Stage 3 Full✓✓✓91.0

Q3c: How does the future prediction loss differ from imitation learning, does it penalize other valid trajectories, and why is Stage 3 considered on-policy adaptation?

Response 3c: We thank the reviewer for the question. We acknowledge that the future prediction loss relies on expert demonstrations. Unlike standard imitation learning, which directly matches expert trajectories, our method supervises planning through future representation prediction. Our objective is to leverage the learned predictive capability to improve planning. We do not aim to address the multimodality of valid driving trajectories, as the reviewer appears to have assumed.

Although Stage 3 uses on-policy rollouts to optimize the planner on its own generated trajectories, the supervision target still comes from expert demonstrations. Consequently, without closed-loop environment feedback, other valid trajectories may still be penalized. Addressing this limitation through closed-loop reinforcement learning is beyond the scope of this work and remains a promising direction for future research.


Q3d: Stage 3 improvements on NAVSIM v1/v2 and NavHard.

Response 3d: We thank the reviewer for this question. We first clarify that all main results reported in Tables 1–3 correspond to the final models after completing all three training stages. To further quantify the contribution of Stage 3, we compare the performance of Stage 2 and Stage 3 on NAVSIM v1, NAVSIM v2, and NavHard.

Table R6. Performance comparison between Stage 2 and Stage 3.

Training StageNAVSIM v1 (PDMS ↑)NAVSIM v2 (EPDMS ↑)NavHard (EPDMS ↑)
Stage 290.890.232.3
Stage 391.090.834.4

Q3e: Different roles of future prediction loss in Stages 2 and 3.

Response 3e: We thank the reviewer for this question. The future prediction loss serves different purposes in Stage 2 and Stage 3, which motivates the different gradient propagation strategies. In Stage 2, it optimizes the encoder and future predictor under GT trajectory conditioning, without updating the planner. This separates representation learning from planning optimization and avoids guiding the planner with an immature future predictor. In Stage 3, the encoder and future predictor are frozen, and the prediction loss optimizes the planner through the learned future predictor under planner-generated trajectory conditioning. Thus, Stage 2 establishes predictive representations, while Stage 3 leverages the learned predictive capability to further improve planning.

AuthorsOpenReview ↗

Response to Reviewer aTwP 3/3

Q3f: Frozen predictor reliability and potential exploitation.

Response 3f: We thank the reviewer for the concern. Fig. 4 shows that trajectories closer to a fixed reference exhibit higher similarity between predicted future representations, suggesting a meaningful local optimization direction. To directly evaluate the actual prediction-loss gradients, we further conduct a one-step update experiment on 1,000 planner-generated trajectories. As shown in Table R7, gradient descent on the future prediction loss improves GT trajectory consistency more frequently than random updates. Together with the Stage 3 loss ablation in Table R5, these results support the effectiveness of the frozen predictor in providing useful planning gradients.

Table R7. Effectiveness of future prediction gradients on planner-generated trajectories.

Update DirectionGT L1 Improvement Ratio ↑
Random Direction46.4%
Emb Gradient Descent70.9%

Q3g: Are the gains from Stage 3 simply due to additional training steps?

Response 3g: We thank the reviewer for the question. Stage 2 is trained for 35K steps and has already converged. As shown in Table R8, extending Stage 2 training does not yield further improvements. In contrast, Stage 3 achieves additional performance gains, demonstrating that its improvements cannot be attributed solely to additional training steps.

Table R8. Performance with extended Stage 2 training.

Stage 2 Training StepsPDMS ↑
20K90.8
25K90.6
30K90.7
35K90.8
+Stage 3 5K steps91.0

Question 4: Comparison with generative world model representations.(W4 & Q5)

Response 4: We replace the V-JEPA2 encoder with the pretrained encoder from Orbis while maintaining the same Stage 2 and Stage 3 training framework. As shown in Table R9, ReDrive consistently improves planning performance with the Orbis encoder, demonstrating its applicability to generative world model representations. Meanwhile, ReDrive achieves better performance across both stages, supporting our choice of pretrained visual backbone.

Table R9. Performance of ReDrive with different pretrained visual encoders.

EncoderStageNC ↑DAC ↑EP ↑C ↑TTC ↑PDMS ↑
OrbisStage 299.297.283.3100.097.290.2
OrbisStage 399.297.684.199.997.290.7
ReDriveStage 299.197.784.4100.096.890.8
ReDriveStage 399.197.984.3100.097.291.0

Question 5: Completeness of benchmark comparisons and inference efficiency. (W5 & Q6)

Q5a: Completeness of method comparisons across NAVSIM v1 and v2.

Response 5a:We thank the reviewer for pointing this out. These methods were not included in the original NAVSIM v2 comparison because their reported results used the previous EPDMS* metric, while our main comparison uses the updated EPDMS metric. For a more comprehensive comparison, we include these methods in Table R10 under the consistent EPDMS* metric, where ReDrive outperforms all additional methods.

Table R10. Additional comparisons on NAVSIM v2 using EPDMS*.

MethodNC ↑DAC ↑DDC ↑TLC ↑EP ↑TTC ↑LK ↑HC ↑EC ↑EPDMS* ↑
TransFuser96.989.997.899.787.195.492.798.387.276.7
VADv298.098.399.499.887.196.895.298.388.185.8
Hydra-MDP++98.897.899.1100.084.095.370.1–96.884.1
DriveSuprim97.897.999.599.990.697.196.698.377.986.0
Drive-JEPA98.897.499.099.883.598.096.298.185.685.4
DAWN97.392.099.199.787.496.696.098.385.583.2
ReDrive99.197.799.699.987.698.798.298.386.587.1

Q5b: Inference latency and parameter count comparisons.

Response 5b: As shown in Table R11, ReDrive achieves competitive inference efficiency on an NVIDIA RTX 4090, with baseline results taken from MORE (2026). While Drive-JEPA achieves lower latency without iterative denoising, ReDrive maintains competitive inference efficiency alongside strong planning performance.

Table R11. Comparison of model parameters and inference latency. All latency measurements are reported on an NVIDIA RTX 4090 GPU with batch size 1.

MethodResolutionFramesCamerasParams (M) ↓Latency (ms) ↓
Drive-JEPA256×51221311.273.3
World4Drive384×6404670.2414.5
Epona512×10241012915.4900.0
DriveLaW768×1344413301.61167.0
CoWorld-VLA512×1024115372.9378.4
ReDrive (NFE=4)256×51241407.0222.4

Comments on the paper

1 comment

Dear Reviewers,

Please take time to review the responses from the authors and engage with them for further questions and comments if any.

Time left is not much and please do shortly.

Thanks

AC

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

Related papers

Discussion (0)

Sign in to comment.