acceptodds
Under review as a conference paper at ICLR 2027

IDOL: Inverse-Dynamics-Guided Future Prediction for End-to-End Autonomous Driving

Abstract

End-to-end autonomous driving has emerged as a compelling paradigm for learning planning directly from sensor observations, while recent world-model-based approaches further enrich this paradigm by enabling explicit reasoning about how the scene may evolve in the future. Yet future prediction alone does not guarantee better planning unless the predicted evolution can be converted into planning-relevant trajectory updates. Many current methods still forecast future scene states without explicitly decoding the motion implications hidden in state transitions. As a result, future reasoning often remains descriptively useful but only weakly coupled to executable motion generation. To address this limitation, we propose **IDOL**, an inverse-dynamics-guided future prediction framework for world-model-based end-to-end planning in latent BEV space, where inverse dynamics serves as the key bridge between future prediction and trajectory optimization. IDOL first predicts multiple future latent scene states with a BEV world model, then applies an inverse dynamics model to adjacent latent futures to decode transition-aware trajectory features and recover planning-relevant motion deltas that explain how the latent world evolves over time. These transition features refine the candidate-specific planning queries through spatial cross-attention and global modulation. A lightweight closed-loop refinement module feeds the refined queries back into candidate-conditioned latent rollouts, enabling an additional round of future-informed query refinement before trajectory decoding. By introducing inverse dynamics into latent future reasoning, IDOL tightens the coupling between world modeling and planning. Extensive experiments on the NAVSIM v1 and NAVSIM v2 benchmarks show that IDOL achieves state-of-the-art performance among comparable methods.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, World Model, Inverse Dynamics, Latent BEV, Trajectory Planning, Future Prediction, Closed-Loop Refinement

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_pfa5OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents IDOL, an end-to-end driving planner that uses transitions between predicted future BEV states to refine trajectory planning. For each trajectory anchor, a BEV world model predicts candidate-conditioned future states. An MLP-based inverse dynamics module encodes adjacent states into spatial and global transition features, which update the planning queries through cross-attention and modulation. The updated queries support another round of future reasoning before trajectory decoding and selection. An auxiliary objective supervises motion corrections relative to the expert trajectory. Experiments cover NAVSIM v1, NAVSIM v2, and Bench2Drive, with particularly strong results on the challenging navhard evaluation.

Critical strengths
  1. Strong performance on challenging driving scenarios, supported by evaluation across benchmarks. The most compelling result is the two-stage NAVSIM v2 navhard evaluation, where IDOL achieves 38.0 EPDMS compared with 27.9 for WoTE. The paper also reports improvements on Bench2Drive, reaching a Driving Score of 63.22 and a success rate of 34.09%, compared with 61.71 and 31.36% for WoTE. These results are consequential to my recommendation: they indicate that the approach is useful under difficult evaluation conditions and in interactive closed-loop simulation.

  2. The component studies provide evidence for the proposed transition representation. I appreciate the parameter-matched comparisons in Table 12. IDOL reaches 90.0 PDMS, compared with 88.9 for future-state features with a matched MLP and 89.1 for latent differences with a matched MLP. Since the remaining components and training settings are held fixed, these results support jointly encoding ordered state pairs. The branch and refinement ablations provide further supporting evidence. The reported inference speed of 17.65 FPS on a single RTX 3090 also suggests that the additional reasoning has a manageable computational cost.

Critical weaknesses

I do not find a weakness that warrants rejection. I have two remaining concerns that would improve the interpretation of the results:

  1. The inverse dynamics interpretation could be better established. The module produces latent transition features, while its explicit motion supervision predicts the difference between expert and anchor trajectories. This is a reasonable way to learn planning-relevant corrections, but it does not directly demonstrate recovery of the motion responsible for the predicted state transition. An ablation of the motion-correction loss or a direct evaluation of the encoded motion information would help clarify this interpretation. This concern does not undermine the demonstrated planning benefit.
  2. The strongest benchmark improvement would benefit from corresponding component analysis. Most component ablations are conducted on navtest, whereas the largest improvement appears in the two-stage navhard setting, which uses the camera-based IDOL-LTF variant. Repeating the main ablations under that configuration would help explain which components account for the improvement in difficult scenarios. I view this as a gap in attribution rather than a reason to discount the reported result.
Questions for authors
  1. How much does the motion-correction loss LcorrL_{\mathrm{corr}} contribute when the rest of the architecture and training setup are unchanged? Do the learned features recover motion information beyond what is already encouraged by the trajectory-planning losses?
  2. Could the authors clarify the checkpoint and input configuration used for each navhard evaluation and provide the main component ablations under the same IDOL-LTF setup? In particular, how much do transition encoding and the additional query-refinement round contribute to the two-stage result?
3 comments
AuthorsOpenReview ↗

Thanks and response to Reviewer pfa5

We thank the reviewer for the insightful assessment and constructive suggestions. We address the remaining concerns below with additional motion-encoding analysis, controlled ablations, and clarification of the IDOL-LTF evaluation protocol.

Q1. How much does the motion-correction loss contribute, and do the learned features encode motion information beyond trajectory-planning supervision?

Correction-loss ablation. We evaluate the contribution of the motion-correction loss while keeping the remaining architecture and training configuration unchanged.

VariantNC ↑DAC ↑TTC ↑Comfort ↑EP ↑PDMS ↑
IDOL w/o L_corr98.997.395.7100.083.789.7
IDOL98.797.795.5100.084.090.0

Removing L_corr reduces PDMS from 90.0 to 89.7, confirming that motion-correction supervision provides an additional benefit beyond the trajectory-planning objectives.

Motion-information analysis. To examine whether IDM encodes motion information beyond expert-relative corrections, we freeze the planner and train linear probes to recover the conditioning anchor's speed and heading rate. Results are evaluated on anchor identities held out from probe fitting.

RepresentationSpeed R² ↑Heading-rate R² ↑
Future-only MLP0.70000.2900
Difference MLP0.73000.3500
Global IDM feature0.76620.4084

The global IDM feature achieves the strongest linear recovery, indicating that it encodes accessible candidate-motion information rather than merely expert-relative corrections. Furthermore, retaining IDM updates reduces predicted-trajectory speed and heading-rate MAE by 11.38% and 15.80%, respectively, compared with bypassing these updates in the frozen planner. Together, these results support the motion-encoding capability and planning benefit of IDM, while L_corr further aligns the learned representation with actionable trajectory refinement.

Q2. What checkpoint and input configuration are used for navhard evaluation, and how do IDM and closed-loop refinement contribute under the same IDOL-LTF setup?

Evaluation configuration. The reported IDOL-LTF navhard result adopts a unified checkpoint for both evaluation stages. Inputs comprise the current front-left, front, and front-right camera images and ego status. The LTF configuration is enabled: the learned latent tensor stored in the checkpoint replaces LiDAR input. We will provide the checkpoint checksum and evaluation configuration for reproducibility.

IDOL-LTF component ablations. We further evaluate the contributions of IDM and closed-loop refinement under the same NAVSIM v2 navhard two-stage protocol. All 5,912 scenarios (450 in Stage 1 and 5,462 in Stage 2) are evaluated successfully.

MethodStageNC ↑DAC ↑DDC ↑TLC ↑EP ↑TTC ↑LK ↑HC ↑EC ↑EPDMS ↑
IDOL w/o IDMStage 196.586.897.299.481.795.193.396.671.2
Stage 284.677.484.997.380.881.848.596.862.829.5
IDOL w/o CLStage 197.186.298.099.681.796.794.997.676.4
Stage 284.880.987.598.184.680.349.796.565.737.2
IDOL (Full)Stage 197.289.698.099.682.396.995.397.672.9
Stage 284.981.787.798.584.781.150.095.464.038.0

Removing closed-loop refinement reduces EPDMS from 38.0 to 37.2 (-0.8), while removing IDM results in a substantially larger drop to 29.5 (-8.5). These results demonstrate that both components contribute to performance under the challenging navhard protocol, with IDM playing a particularly important role in maintaining robust planning performance. This further supports the effectiveness of transition-aware planning and iterative refinement in the camera-based IDOL-LTF configuration.

We sincerely thank the reviewer again for the constructive feedback and valuable suggestions. We will promptly incorporate the additional experiments, analyses, and clarifications into the revised manuscript to further strengthen the paper.

Thank you for the additional analyses. The correction-loss ablation and motion probes help clarify what the transition features encode. The navhard ablations are particularly useful: the drop from 38.0 to 29.5 without IDM supports its contribution in the challenging setting, while the smaller drop without refinement clarifies the relative roles of the two components.

These results largely address my remaining questions. They strengthen the evidence for useful motion-aware representations, although I would still distinguish this from demonstrating inverse dynamics in the conventional sense. This distinction does not change my positive assessment of the planning results.

Please incorporate these analyses and the evaluation configuration into the manuscript. Please also reconcile the stage-wise EPDMS values of 76.2/48.8 in the response to WSWz with the 38.0 result in the ablation table, so the reporting is unambiguous.

I remain supportive of acceptance and maintain my rating of 4 (clear acceptance).

AuthorsOpenReview ↗

Thanks for the feedback

We sincerely appreciate Reviewer pfa5’s positive assessment, continued support, and constructive feedback. We will incorporate the additional analyses and evaluation details into the revised manuscript and explicitly distinguish the stage-wise EPDMS scores (76.2/48.8) from the final two-stage aggregated score (38.0) to ensure clear and consistent reporting.

Reviewer_ina2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Unlike conventional approaches that directly leverage predicted latent bev representations to guide end-to-end autonomous driving planners, this paper introduces an inverse dynamics model to extract motion-relevant information from predicted latent bev state transitions, enabling more effective trajectory refinement. Experimental results on NAVSIM v1 and v2 demonstrate strong performance compared with baseline approaches, particularly in challenging driving scenarios.

Critical strengths
  • The proposed idea is reasonable and well-motivated, providing an intuitive way to leverage predicted future dynamics for trajectory refinement.
  • The ablation studies demonstrate the effectiveness of the proposed approach and validate the contributions of its key components.
Critical weaknesses
  • My major concern is the lack of comparisons with recent sota vla and wam approaches that have demonstrated stronger performance. Furthermore, the paper does not investigate whether the proposed method can be integrated into these approaches to achieve further performance improvements. Such experiments would significantly strengthen the paper's claims and provide stronger evidence of the proposed method's effectiveness and generalizability.
  • The paper reports the overall inference speed, but lacks a detailed analysis of the additional computational cost introduced by the inverse dynamics model and iterative refinement.
Questions for authors
  • I would like to see experiments that integrate the proposed method into recent sota vla / wam approaches to demonstrate whether it can achieve further performance improvements.
  • Please provide a detailed analysis of the additional computational cost introduced by the inverse dynamics model and iterative refinement
2 comments
AuthorsOpenReview ↗

Thanks and response to Reviewer ina2

We thank the reviewer for the constructive comments and valuable suggestions. We agree that comparisons with recent VLA/WAM approaches and cross-architecture validation are important for demonstrating the effectiveness and generalizability of our method. Accordingly, we have added such comparisons and cross-architecture validation results, together with a computational-cost analysis, as detailed below.

Q1. How does IDOL compare with recent VLA/WAM approaches, and can it generalize to other planning frameworks?

Comparison with recent VLA/WAM approaches. We further compare IDOL with four representative methods: SGDrive-IL [1], which introduces hierarchical scene-agent-goal reasoning into VLA-based planning; DriveLaW [2], which unifies video generation and motion planning through shared latent representations; WorldDrive [3], which integrates trajectory-aware world modeling with future-aware planning; and ReCogDrive-IL [4], which combines VLM-based cognitive reasoning with diffusion-based trajectory planning. These methods represent complementary VLA and world-model-based planning paradigms.

The quantitative comparisons on NAVSIM v1 navtest are summarized below.

MethodParams (M) ↓NC ↑DAC ↑TTC ↑Comfort ↑EP ↑PDMS ↑
SGDrive-IL [1]~2,00098.695.195.4100.081.287.4
DriveLaW [2]3,301.699.097.196.7100.081.389.1
WorldDrive [3]N/R98.496.895.2100.083.389.0
ReCogDrive-IL [4]~2,03598.194.794.2100.080.986.5
IDOL (Ours)69.3698.797.795.5100.084.090.0

N/R denotes not reported. SGDrive-IL and ReCogDrive-IL refer to their imitation-learning configurations. WorldDrive uses the full-navtrain configuration reported in its paper.

Our IDOL achieves 90.0 PDMS, outperforming all four selected configurations while maintaining a compact architecture with only 69.36M parameters. In particular, IDOL achieves the highest drivable area compliance (97.7%) and ego progress (84.0%) among these methods. Although differences in backbone architectures, sensor configurations, and training strategies preclude strictly controlled comparisons, these results demonstrate the competitiveness and parameter efficiency of our approach.

Integration into Latent-WAM. To further validate the generalizability of our IDM, we integrate it into Latent-WAM [5], a lightweight world-action model that learns compact latent world representations. We adapt our IDM to extract motion-aware features from consecutive latent states during training. Through motion supervision and feature distillation, a lightweight student branch learns to predict these dynamics from current-frame representations. The resulting features are incorporated into Latent-WAM's trajectory decoder to enhance planning, while preserving its original current-frame-only inference setting without requiring future observations or additional world-model rollouts.

The experimental results on NAVSIM v2 navtest are summarized below.

MethodNC ↑DAC ↑DDC ↑TLC ↑EP ↑TTC ↑LK ↑HC ↑EC ↑EPDMS ↑
Latent-WAM [5]98.197.399.699.887.797.397.698.187.389.3
Latent-WAM + IDM (Ours)98.797.699.199.888.798.196.598.385.189.8

Integrating our IDM improves Latent-WAM's EPDMS from 89.3 to 89.8 (+0.5 points), with notable improvements in ego progress (+1.0 points) and time-to-collision (+0.8 points). These results demonstrate that inverse-dynamics-guided representation learning can benefit a different world-action modeling architecture, providing further evidence of the transferability and generalizability of our approach.

[1] SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving, CVPR 2026.

[2] DriveLaW: Unifying Planning and Video Generation in a Latent Driving World, CVPR 2026.

[3] Bridging Scene Generation and Planning: Driving with World Model via Unifying Vision and Motion Representation, NeurIPS 2026.

[4] ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving, ICLR 2026.

[5] Latent-WAM: Latent World Action Modeling for End-to-End Autonomous Driving, ACM MM 2026.

AuthorsOpenReview ↗

Thanks and response to Reviewer ina2

Q2. What is the additional computational cost introduced by the IDM and iterative refinement?

We conduct a controlled analysis to isolate the computational overhead introduced by the inverse dynamics model (IDM) and iterative refinement. All measurements are conducted on an RTX 3090 GPU with batch size 1, identical inputs, 100 timed inference runs, and the same checkpoint.

Inference SettingLatency / ScenarioFLOPs / ScenarioFPS
w/o IDM31.03 ms105.55 GFLOPs32.22 scenarios/s
IDM31.64 ms125.25 GFLOPs31.60 scenarios/s
IDM+closed-loop refinement46.30 ms182.32 GFLOPs21.60 scenarios/s

The 21.60 scenarios/s reported here is the reciprocal of model inference latency under our controlled timing protocol; the 17.65 FPS reported in the paper uses end-to-end timing.

Enabling IDM introduces only a modest computational overhead, increasing latency by 0.61 ms and FLOPs by 19.70 GFLOPs ​. In contrast, the additional closed-loop refinement pass incurs 14.65 ms and ​57.07 GFLOPs ​, primarily due to repeated world-model inference for candidate trajectory refinement. Overall, the full two-pass configuration increases latency by 15.26 ms compared with the baseline without IDM, while maintaining a throughput of ​21.60 scenarios/s*. These results indicate that the computational overhead is mainly attributable to iterative refinement rather than the IDM itself.

We sincerely thank the reviewer again for the constructive feedback and valuable suggestions. We will promptly incorporate the additional experiments, analyses, and clarifications into the revised manuscript to further strengthen the paper.

Reviewer_WSWzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes IDOL, an end-to-end planner for NAVSIM that adds an "inverse dynamics model" (IDM) to a WoTE(Li et al.,2025e) style world-model planner. The pipeline consists of a TransFuser (Chitta et al., 2022)-style encoder (ResNet-34, camera + LiDAR) that produces a latent BEV map. A transformer world model rolls the latent BEV forward, with that query injected at the anchor's future location. A lightweight MLP reads each pair of adjacent imagined states, concatenated along the channels. It outputs a spatial dynamics map and a pooled global feature, and both are averaged over the rollout steps. Cross-attention over the spatial map and MLN modulation by the global feature refine the query. The refined query seeds one more anchor-conditioned rollout ("closed-loop refinement", Eq. 11). Offsets are decoded after the last pass, and WoTE's reward model ranks the candidates. A training-only head on the final global IDM feature of the anchor closest to the expert regresses the expert-minus-anchor speed, acceleration and heading-rate sequences.

Critical strengths

A marginal gain on SOTA has been demonstrated.

It addresses a key question: for the NAVSIM world-model line (LAW, WoTE, SeerDrive, ResWorld), how should a world-model planner turn predicted futures into trajectory changes rather than extra context? The inverse dynamics model (IDM adds only about 1.2M parameters.

IDOL is evaluated on NAVSIM v1 navtest, NAVSIM v2 navtest, navhard stage-1, two-stage navhard and closed-loop Bench2Drive. It improves on WoTE, whose rollouts and reward model it builds on, in every table where WoTE appears

Critical weaknesses

W1. The paper does not show that the module performs inverse dynamics. Each rollout injects the planning query at the anchor's future location. The motion that "explains" an imagined transition is therefore the anchor's displacement, which the planner already has. Whatever else the module extracts, for example, how other agents move relative to the anchor, makes it a learned critic of the candidate rather than an inverse model.

W2. IDM L_corr (weight 0.2) is the only direct supervision of the IDM. It applies to the anchor closest to the expert at the final pass, and its target is a correction toward the expert, not the action that takes State_t to State_t+1 in Eq. 1.

W3. The state-of-the-art claims hold only within the ResNet-34 class.

W4. The no-IDM base is weaker than the published system it resembles. Table 3's no-IDM, no-refinement model scores 87.3 PDMS. That is below the published WoTE (88.3, Table 1), whose reward model IDOL reuses, and below Table 12's future-state-only variant (88.6). The paper does not define what this row feeds the planner. Measured against published WoTE, the IDM adds 0.9 PDMS and the full model 1.7, rather than the 1.9 and 2.7 implied in Section 4.4. Against the capacity-matched future-state control, IDOL adds 1.1.

W5. The introduction says the module "re-evaluates the updated trajectory", but Section 3.4.2 keeps the anchor fixed and decodes offsets only after the final pass. The world model therefore never imagines a refined trajectory; the second rollout re-processes the same anchor with an updated query. "Long-horizon consistency" is never measured, the re-anchoring of Eq. 11 is not ablated, and no latency is reported with and without the extra rollout of all 256 candidates.

Questions for authors

What does the IDM encode? What happens when Loss_corr = 0, and do the Table 12 controls receive an equivalent correction loss? Can a linear probe on the global IDM feature recover the conditioning anchor's speed and heading rate, rather than the expert correction, and how large is the Loss_corr error on navtest?

What happens at test time if you swap the order of each pair, or pair states from different anchors' rollouts, and what share of the IDM response falls on agent or route cells over navtest?

Two-stage navhard. What are the per-stage EPDMS values and scenario counts for every Table 2 row? Which NAVSIM version and navhard metric (before or after the fix that DrivoR describes) do your rows and the copied rows use, and which rows were re-run rather than copied? What does IDOL-LTF score on navtest, and what does it score on the two-stage protocol without the IDM?

What does the no-IDM, no-refinement (CL iters) row of Table 3 feed the planner, and which ranker does it use? Why does it score 87.3 when the published WoTE scores 88.3 and Table 12's future-state-only variant scores 88.6; is it WoTE retrained with your schedule?

2 comments
AuthorsOpenReview ↗

Thanks and response to Reviewer WSWz

We thank the reviewer for the detailed questions and helpful suggestions. We address them below with additional motion probes, ablations, and evaluation details.

Q1. What does the IDM encode, and what is the role of correction supervision?

Our IDM provides an explicit bridge between predicted states and trajectory planning: it converts candidate-conditioned state transitions into a representation used to refine planning queries. Rather than relying solely on future-state features, this design makes transition information directly available to the trajectory-refinement pathway. This is the planning-oriented role of inverse dynamics in IDOL.

Linear recovery of anchor motion. Following your suggestion, we freeze the planner and fit linear probes to recover the conditioning anchor's speed and heading rate, rather than the expert correction. The probes use the same evaluation configuration; the table reports results on anchor identities held out from probe fitting.

MethodSpeed R² ↑Heading-rate R² ↑
Future-only MLP0.70000.2900
Difference MLP0.73000.3500
Global IDM feature0.76620.4084

The global IDM feature supports stronger linear recovery than these two MLP controls, demonstrating accessible candidate-motion information. Separately, retaining the IDM updates reduces predicted-trajectory speed and heading-rate MAE by 11.38% and 15.80%, respectively, compared with bypassing the updates in the frozen planner. These results support both motion encoding and useful trajectory refinement.

Correction supervision. In our paper, all Table 12 variants receive equivalent correction supervision. ​We conduct an ablation study on the correction loss.​​ The correction-loss ablation gives:

MethodNC ↑DAC ↑TTC ↑Comfort ↑EP ↑PDMS ↑
IDOL w/o L_corr98.997.395.7100.083.789.7
IDOL98.797.795.5100.084.090.0

Inverse dynamics provides the planning-relevant transition representation, while motion-correction supervision further aligns this representation with actionable trajectory refinement. The auxiliary loss supervises expert-relative motion correction, while the linear-probe targets above are the conditioning anchor's own motion. The reported training and validation correction losses are 8.36e-4 and 1.05e-3, respectively, under the training loss's normalization.

Q2. What happens when state pairs are perturbed, and where is the IDM response concentrated?

We test the frozen planner on 512 navtest scenes spanning 136 logs, perturbing the pair at each IDM invocation while retaining the candidate selected by the unmodified model.

MethodADE (m) ↓FDE (m) ↓
Matched pair0.68521.6840
Reversed order0.74821.7952
Different anchors0.75161.8005

Reversing order and mixing anchors increase ADE by 0.0630 m and 0.0664 m, respectively; both paired log-bootstrap 95% confidence intervals exclude zero. This demonstrates that ordered, candidate-matched transitions matter for refinement.

For the final IDM pass and selected candidate, we measure attention-weighted response using fractional BEV-cell occupancy under the semantic-grid mapping:

RegionArea shareAttention-weighted IDM responseResponse / area
Agents2.84%4.19%1.48×
Road35.29%53.50%1.52×

The response is enriched over driving-relevant regions relative to their area. Here, road cells cover lanes and intersections, excluding agent-covered portions.

AuthorsOpenReview ↗

Thanks and response to Reviewer WSWz

Q3. Two-stage navhard protocol, provenance, and IDOL-LTF results.

We further evaluate the contributions of IDM and closed-loop refinement under the NAVSIM v2 navhard two-stage protocol. All 5,912 scenarios (450 in Stage 1 and 5,462 in Stage 2) are evaluated successfully. The full IDOL-LTF model achieves an ​EPDMS of 76.2 in Stage 1 and 48.8 in Stage 2​. The full IDOL-LTF model and two ablated variants yield the following results:

MethodStageNC ↑DAC ↑DDC ↑TLC ↑EP ↑TTC ↑LK ↑HC ↑EC ↑EPDMS ↑
IDOL w/o IDMStage 196.586.897.299.481.795.193.396.671.2
Stage 284.677.484.997.380.881.848.596.862.829.5
IDOL w/o CLStage 197.186.298.099.681.796.794.997.676.4
Stage 284.880.987.598.184.680.349.796.565.737.2
IDOL (Full)Stage 197.289.698.099.682.396.995.397.672.9
Stage 284.981.787.798.584.781.150.095.464.038.0

Removing closed-loop refinement reduces EPDMS from 38.0 to 37.2 (-0.8), while removing IDM leads to a substantially larger drop to 29.5 (-8.5). These results confirm that both components contribute to two-stage planning performance, with IDM providing the larger gain under challenging navhard scenarios.

On navtest, the same full IDOL-LTF navhard checkpoint achieves 88.1 EPDMS:

MethodNCDACDDCTLCEPTTCLKHCECEPDMS
IDOL-LTF98.4596.4199.3599.8386.8897.8897.0498.3085.5288.1

Our IDOL results are from our own evaluations. Some comparison results are cited from publicly available works, including baseline or reproduction results reported by GuideFlow [1] and MindDrive [2]. The cited PDM-Closed and LTF results use the pre-fix metric.

Q4. What does the 87.3 baseline contain, and how does it differ from WoTE and the 88.6 variant?

In the no-IDM/no-CL row, the trajectory decoder takes the initial anchor-and-ego-state queries and current BEV features, predicting offsets added to the anchors without IDM updates or iterative feedback. The world model still performs one rollout, and candidates are ranked using the same ranker architecture and scoring rule as the full model.

This is an internal IDOL ablation, with implementation and training differences from WoTE: IDOL omits WoTE's preliminary current-scene-based trajectory-offset prediction before rollout, the additional inference-time Offset Anchor re-encoding, and Agent Detection Loss, among other design elements. Its remaining final offset decoder is the one described above. The 88.6 future-state-only variant additionally introduces predicted future features into the planning pathway. These configurations therefore have different information pathways and are not identical WoTE reproductions.

We sincerely thank the reviewer again for the constructive feedback and valuable suggestions. We will promptly incorporate the additional experiments, analyses, and clarifications into the revised manuscript to further strengthen the paper.

[1] GuideFlow: Constraint-Guided Flow Matching for Planning in End-to-End Autonomous Driving, CVPR 2026.

[2] MindDrive: An All-in-One Framework Bridging World Models and Vision-Language Model for End-to-End Autonomous Driving, arXiv:2512.04441, 2025.

Comments on the paper

1 comment
Area_Chair_VqWeOpenReview ↗

Please Review the Author Rebuttal and Reassess Your Rating

Dear Reviewers,

Thank you for your time and effort in reviewing the submissions.

The authors have now submitted their rebuttals. I kindly ask you to carefully review the authors' responses and assess whether your questions and concerns have been adequately addressed.

If the rebuttal resolves your concerns, please consider updating your review and adjusting your rating accordingly. If any concerns remain, please feel free to raise them in the discussion.

Your timely feedback and active participation are greatly appreciated, as they are essential to reaching a fair, well-informed decision.

Thank you again for your valuable contributions!

Best regards,
Area Chair
ICLR

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.